Model profile · OpenAI

GPT-5.4

Closed weightsfrontier · registry tag 2026 flagship

Verified but aging

Reads as verified but aging across the resolved source data. Capability fit excludes speed and cost.

Visible coverage: 66.4%
Verified coverage: 64.4%
Spread: 87.5%
Last verified: Oct 3, 2026

69%capability fit

textcodevisiondocumentsearch19 aliases53 official source links

Open compare

Data version

Current snapshot.

Published Oct 3, 2026Model list checked4 stale sources105 claim warnings9 providers · 1334 tracked models

The registry snapshot and page stamp are shown so a stale deploy is visible at a glance. View live data status.

Source-linked scores by benchmark

Each row keeps the benchmark source, source type, raw metric, and percentile inside its fair comparison set. Capability summaries exclude archived categories, speed, and cost; those rows remain below.

Verified but agingThis model currently reads as verified but aging across the resolved source data.

Chat / text36 benchmarks69.8%

Intelligence Index

AA · Chat / text · Combined

It tests whether the model is actually useful in normal conversational turns, not just on narrow correctness tasks.

Rank #182 · Source label: GPT-5.4 (Non-reasoning)

verified runtimeexact alias

Raw row drilldownsource row, percentile, last updated, eligibility

Source: Artificial Analysis
Raw value: 18
Percentile: 67.4%
Last updated: recent
Eligibility: headline eligible

Parsed from Artificial Analysis public leaderboard field `intelligenceIndex`.

67.4% percentile inside its fair comparison set

18Raw benchmark value

AA-Omniscience accuracy

AA · Chat / text · Objective

It tests whether the model is actually useful in normal conversational turns, not just on narrow correctness tasks.

Rank #97 · Source label: GPT-5.4 (Non-reasoning)

verified runtimeexact alias

Raw row drilldownsource row, percentile, last updated, eligibility

Source: Artificial Analysis
Raw value: 37.4%
Percentile: 78.8%
Last updated: recent
Eligibility: headline eligible

Parsed from Artificial Analysis public leaderboard field `omniscienceAccuracy`.

78.8% percentile inside its fair comparison set

37.4%Raw benchmark value

AA-Omniscience non-hallucination

AA · Chat / text · Objective

It tests whether the model is actually useful in normal conversational turns, not just on narrow correctness tasks.

Rank #364 · Source label: GPT-5.4 (Xhigh)

verified runtimeexact alias

Raw row drilldownsource row, percentile, last updated, eligibility

Source: Artificial Analysis
Raw value: 8.3%
Percentile: 19.9%
Last updated: recent
Eligibility: headline eligible

Parsed from Artificial Analysis public leaderboard field `omniscienceNonHallucination`.

19.9% percentile inside its fair comparison set

8.3%Raw benchmark value

IFBench

AA · Chat / text · Objective

It tests whether the model is actually useful in normal conversational turns, not just on narrow correctness tasks.

Rank #116 · Source label: GPT-5.4 (Non-reasoning)

verified runtimeexact alias

Raw row drilldownsource row, percentile, last updated, eligibility

Source: Artificial Analysis
Raw value: 48.4%
Percentile: 67%
Last updated: recent
Eligibility: headline eligible

Parsed from Artificial Analysis public leaderboard field `ifbench`.

67% percentile inside its fair comparison set

48.4%Raw benchmark value

Input price

AA · Chat / text · Speed / cost

It tests whether the model is actually useful in normal conversational turns, not just on narrow correctness tasks.

Rank #304 · Source label: GPT-5.4 (Xhigh)

verified runtimeexact alias

Raw row drilldownsource row, percentile, last updated, eligibility

Source: Artificial Analysis
Raw value: $2.5 /1M input tokens
Percentile: 14.7%
Last updated: recent
Eligibility: headline eligible

Parsed from Artificial Analysis public leaderboard field `price1mInputTokens`.

14.7% percentile inside its fair comparison set

$2.5 /1M input tokensRaw benchmark value

Output price

AA · Chat / text · Speed / cost

It tests whether the model is actually useful in normal conversational turns, not just on narrow correctness tasks.

Rank #309 · Source label: GPT-5.4 (Xhigh)

verified runtimeexact alias

Raw row drilldownsource row, percentile, last updated, eligibility

Source: Artificial Analysis
Raw value: $15 /1M output tokens
Percentile: 13.8%
Last updated: recent
Eligibility: headline eligible

Parsed from Artificial Analysis public leaderboard field `price1mOutputTokens`.

13.8% percentile inside its fair comparison set

$15 /1M output tokensRaw benchmark value

Output Speed

AA · Chat / text · Speed / cost

It tests whether the model is actually useful in normal conversational turns, not just on narrow correctness tasks.

Rank #190 · Source label: GPT-5.4 (Low)

verified runtimeexact alias

Raw row drilldownsource row, percentile, last updated, eligibility

Source: Artificial Analysis
Raw value: 59.7 tokens/s
Percentile: 27.9%
Last updated: recent
Eligibility: headline eligible

Parsed from Artificial Analysis public leaderboard field `medianOutputTokensPerSecond`.

27.9% percentile inside its fair comparison set

59.7 tokens/sRaw benchmark value

Time to first token

AA · Chat / text · Speed / cost

It tests whether the model is actually useful in normal conversational turns, not just on narrow correctness tasks.

Rank #260 · Source label: GPT-5.4 (Xhigh)

verified runtimeexact alias

Raw row drilldownsource row, percentile, last updated, eligibility

Source: Artificial Analysis
Raw value: 192.94s
Percentile: 1.9%
Last updated: recent
Eligibility: headline eligible

Parsed from Artificial Analysis public leaderboard field `medianTimeToFirstTokenSeconds`.

1.9% percentile inside its fair comparison set

192.94sRaw benchmark value

Time to first answer token

AA · Chat / text · Speed / cost

It tests whether the model is actually useful in normal conversational turns, not just on narrow correctness tasks.

Rank #258 · Source label: GPT-5.4 (Xhigh)

verified runtimeexact alias

Raw row drilldownsource row, percentile, last updated, eligibility

Source: Artificial Analysis
Raw value: 192.94s
Percentile: 1.9%
Last updated: recent
Eligibility: headline eligible

Parsed from Artificial Analysis public leaderboard field `medianTimeToFirstAnswerTokenSeconds`.

1.9% percentile inside its fair comparison set

192.94sRaw benchmark value

Text Arena

AR · Chat / text · Human

It tests whether the model is actually useful in normal conversational turns, not just on narrow correctness tasks.

Rank #32 · Source label: gpt-5.4-high

verified runtimeexact alias

Raw row drilldownsource row, percentile, last updated, eligibility

Source: Arena
Raw value: 1,475
Percentile: 91.8%
Last updated: recent
Eligibility: headline eligible

Parsed from Arena leaderboard dataset row `gpt-5.4-high`. Category: overall. Source rank: #34. Votes: 64642. Organization: openai. License: Proprietary.

91.8% percentile inside its fair comparison set

1,475Raw benchmark valueCI 1,471 - 1,479

Text Arena · Creative Writing

AR · Chat / text · Human

It tests whether the model is actually useful in normal conversational turns, not just on narrow correctness tasks.

Rank #50 · Source label: gpt-5.4-high

verified runtimeexact alias

Raw row drilldownsource row, percentile, last updated, eligibility

Source: Arena
Raw value: 1,442
Percentile: 86.9%
Last updated: recent
Eligibility: headline eligible

Parsed from Arena leaderboard dataset row `gpt-5.4-high`. Category: creative_writing. Source rank: #55. Votes: 11109. Organization: openai. License: Proprietary.

86.9% percentile inside its fair comparison set

1,442Raw benchmark valueCI 1,436 - 1,449

Text Arena · English

AR · Chat / text · Human

It tests whether the model is actually useful in normal conversational turns, not just on narrow correctness tasks.

Rank #44 · Source label: gpt-5.4-high

verified runtimeexact alias

Raw row drilldownsource row, percentile, last updated, eligibility

Source: Arena
Raw value: 1,476
Percentile: 88.6%
Last updated: recent
Eligibility: headline eligible

Parsed from Arena leaderboard dataset row `gpt-5.4-high`. Category: english. Source rank: #47. Votes: 29222. Organization: openai. License: Proprietary.

88.6% percentile inside its fair comparison set

1,476Raw benchmark valueCI 1,471 - 1,481

Text Arena · Exclude Ties

AR · Chat / text · Human

It tests whether the model is actually useful in normal conversational turns, not just on narrow correctness tasks.

Rank #36 · Source label: gpt-5.4-high

verified runtimeexact alias

Raw row drilldownsource row, percentile, last updated, eligibility

Source: Arena
Raw value: 1,484
Percentile: 90.7%
Last updated: recent
Eligibility: headline eligible

Parsed from Arena leaderboard dataset row `gpt-5.4-high`. Category: exclude_ties. Source rank: #38. Votes: 49582. Organization: openai. License: Proprietary.

90.7% percentile inside its fair comparison set

1,484Raw benchmark valueCI 1,479 - 1,489

Text Arena · Hard Prompts

AR · Chat / text · Human

It tests whether the model is actually useful in normal conversational turns, not just on narrow correctness tasks.

Rank #36 · Source label: gpt-5.4-high

verified runtimeexact alias

Raw row drilldownsource row, percentile, last updated, eligibility

Source: Arena
Raw value: 1,497
Percentile: 90.7%
Last updated: recent
Eligibility: headline eligible

Parsed from Arena leaderboard dataset row `gpt-5.4-high`. Category: hard_prompts. Source rank: #37. Votes: 42068. Organization: openai. License: Proprietary.

90.7% percentile inside its fair comparison set

1,497Raw benchmark valueCI 1,492 - 1,501

Text Arena · Hard Prompts English

AR · Chat / text · Human

It tests whether the model is actually useful in normal conversational turns, not just on narrow correctness tasks.

Rank #45 · Source label: gpt-5.4-high

verified runtimeexact alias

Raw row drilldownsource row, percentile, last updated, eligibility

Source: Arena
Raw value: 1,493
Percentile: 88.2%
Last updated: recent
Eligibility: headline eligible

Parsed from Arena leaderboard dataset row `gpt-5.4-high`. Category: hard_prompts_english. Source rank: #46. Votes: 19665. Organization: openai. License: Proprietary.

88.2% percentile inside its fair comparison set

1,493Raw benchmark valueCI 1,487 - 1,499

Text Arena · Instruction Following

AR · Chat / text · Human

It tests whether the model is actually useful in normal conversational turns, not just on narrow correctness tasks.

Rank #35 · Source label: gpt-5.4-high

verified runtimeexact alias

Raw row drilldownsource row, percentile, last updated, eligibility

Source: Arena
Raw value: 1,471
Percentile: 91%
Last updated: recent
Eligibility: headline eligible

Parsed from Arena leaderboard dataset row `gpt-5.4-high`. Category: instruction_following. Source rank: #36. Votes: 21838. Organization: openai. License: Proprietary.

91% percentile inside its fair comparison set

1,471Raw benchmark valueCI 1,466 - 1,477

Text Arena · Longer Query

AR · Chat / text · Human

It tests whether the model is actually useful in normal conversational turns, not just on narrow correctness tasks.

Rank #42 · Source label: gpt-5.4-high

verified runtimeexact alias

Raw row drilldownsource row, percentile, last updated, eligibility

Source: Arena
Raw value: 1,481
Percentile: 88.5%
Last updated: recent
Eligibility: headline eligible

Parsed from Arena leaderboard dataset row `gpt-5.4-high`. Category: longer_query. Source rank: #43. Votes: 28402. Organization: openai. License: Proprietary.

88.5% percentile inside its fair comparison set

1,481Raw benchmark valueCI 1,476 - 1,487

Text Arena · Multi Turn

AR · Chat / text · Human

It tests whether the model is actually useful in normal conversational turns, not just on narrow correctness tasks.

Rank #18 · Source label: gpt-5.4-high

verified runtimeexact alias

Raw row drilldownsource row, percentile, last updated, eligibility

Source: Arena
Raw value: 1,493
Percentile: 95.5%
Last updated: recent
Eligibility: headline eligible

Parsed from Arena leaderboard dataset row `gpt-5.4-high`. Category: multi_turn. Source rank: #18. Votes: 11988. Organization: openai. License: Proprietary.

95.5% percentile inside its fair comparison set

1,493Raw benchmark valueCI 1,486 - 1,500

Text Arena · No Style Control

AR · Chat / text · Human

It tests whether the model is actually useful in normal conversational turns, not just on narrow correctness tasks.

Rank #30 · Source label: gpt-5.4-high

verified runtimeexact alias

Raw row drilldownsource row, percentile, last updated, eligibility

Source: Arena
Raw value: 1,469
Percentile: 92.3%
Last updated: recent
Eligibility: headline eligible

Parsed from Arena leaderboard dataset row `gpt-5.4-high`. Category: overall. Source rank: #31. Votes: 64642. Organization: openai. License: Proprietary.

92.3% percentile inside its fair comparison set

1,469Raw benchmark valueCI 1,465 - 1,473

Text Arena · Creative Writing · No Style Control

AR · Chat / text · Human

It tests whether the model is actually useful in normal conversational turns, not just on narrow correctness tasks.

Rank #43 · Source label: gpt-5.4-high

verified runtimeexact alias

Raw row drilldownsource row, percentile, last updated, eligibility

Source: Arena
Raw value: 1,439
Percentile: 88.8%
Last updated: recent
Eligibility: headline eligible

Parsed from Arena leaderboard dataset row `gpt-5.4-high`. Category: creative_writing. Source rank: #45. Votes: 11109. Organization: openai. License: Proprietary.

88.8% percentile inside its fair comparison set

1,439Raw benchmark valueCI 1,432 - 1,446

Text Arena · English · No Style Control

AR · Chat / text · Human

It tests whether the model is actually useful in normal conversational turns, not just on narrow correctness tasks.

Rank #39 · Source label: gpt-5.4-high

verified runtimeexact alias

Raw row drilldownsource row, percentile, last updated, eligibility

Source: Arena
Raw value: 1,467
Percentile: 89.9%
Last updated: recent
Eligibility: headline eligible

Parsed from Arena leaderboard dataset row `gpt-5.4-high`. Category: english. Source rank: #41. Votes: 29222. Organization: openai. License: Proprietary.

89.9% percentile inside its fair comparison set

1,467Raw benchmark valueCI 1,462 - 1,472

Text Arena · Exclude Ties · No Style Control

AR · Chat / text · Human

It tests whether the model is actually useful in normal conversational turns, not just on narrow correctness tasks.

Rank #31 · Source label: gpt-5.4-high

verified runtimeexact alias

Raw row drilldownsource row, percentile, last updated, eligibility

Source: Arena
Raw value: 1,473
Percentile: 92%
Last updated: recent
Eligibility: headline eligible

Parsed from Arena leaderboard dataset row `gpt-5.4-high`. Category: exclude_ties. Source rank: #32. Votes: 49582. Organization: openai. License: Proprietary.

92% percentile inside its fair comparison set

1,473Raw benchmark valueCI 1,469 - 1,478

Text Arena · Hard Prompts · No Style Control

AR · Chat / text · Human

It tests whether the model is actually useful in normal conversational turns, not just on narrow correctness tasks.

Rank #26 · Source label: gpt-5.4-high

verified runtimeexact alias

Raw row drilldownsource row, percentile, last updated, eligibility

Source: Arena
Raw value: 1,485
Percentile: 93.4%
Last updated: recent
Eligibility: headline eligible

Parsed from Arena leaderboard dataset row `gpt-5.4-high`. Category: hard_prompts. Source rank: #27. Votes: 42068. Organization: openai. License: Proprietary.

93.4% percentile inside its fair comparison set

1,485Raw benchmark valueCI 1,481 - 1,490

Text Arena · Hard Prompts English · No Style Control

AR · Chat / text · Human

It tests whether the model is actually useful in normal conversational turns, not just on narrow correctness tasks.

Rank #36 · Source label: gpt-5.4-high

verified runtimeexact alias

Raw row drilldownsource row, percentile, last updated, eligibility

Source: Arena
Raw value: 1,479
Percentile: 90.6%
Last updated: recent
Eligibility: headline eligible

Parsed from Arena leaderboard dataset row `gpt-5.4-high`. Category: hard_prompts_english. Source rank: #37. Votes: 19665. Organization: openai. License: Proprietary.

90.6% percentile inside its fair comparison set

1,479Raw benchmark valueCI 1,473 - 1,484

Text Arena · Instruction Following · No Style Control

AR · Chat / text · Human

It tests whether the model is actually useful in normal conversational turns, not just on narrow correctness tasks.

Rank #28 · Source label: gpt-5.4-high

verified runtimeexact alias

Raw row drilldownsource row, percentile, last updated, eligibility

Source: Arena
Raw value: 1,469
Percentile: 92.8%
Last updated: recent
Eligibility: headline eligible

Parsed from Arena leaderboard dataset row `gpt-5.4-high`. Category: instruction_following. Source rank: #29. Votes: 21838. Organization: openai. License: Proprietary.

92.8% percentile inside its fair comparison set

1,469Raw benchmark valueCI 1,464 - 1,475

Text Arena · Longer Query · No Style Control

AR · Chat / text · Human

It tests whether the model is actually useful in normal conversational turns, not just on narrow correctness tasks.

Rank #38 · Source label: gpt-5.4-high

verified runtimeexact alias

Raw row drilldownsource row, percentile, last updated, eligibility

Source: Arena
Raw value: 1,474
Percentile: 89.6%
Last updated: recent
Eligibility: headline eligible

Parsed from Arena leaderboard dataset row `gpt-5.4-high`. Category: longer_query. Source rank: #40. Votes: 28402. Organization: openai. License: Proprietary.

89.6% percentile inside its fair comparison set

1,474Raw benchmark valueCI 1,468 - 1,479

Text Arena · Multi Turn · No Style Control

AR · Chat / text · Human

It tests whether the model is actually useful in normal conversational turns, not just on narrow correctness tasks.

Rank #20 · Source label: gpt-5.4-high

verified runtimeexact alias

Raw row drilldownsource row, percentile, last updated, eligibility

Source: Arena
Raw value: 1,482
Percentile: 94.9%
Last updated: recent
Eligibility: headline eligible

Parsed from Arena leaderboard dataset row `gpt-5.4-high`. Category: multi_turn. Source rank: #20. Votes: 11988. Organization: openai. License: Proprietary.

94.9% percentile inside its fair comparison set

1,482Raw benchmark valueCI 1,475 - 1,489

Instruction following

LB · Chat / text · Objective

It tests whether the model is actually useful in normal conversational turns, not just on narrow correctness tasks.

Rank #29 · Source label: gpt-5.4-xhigh

verified runtimeexact alias

Raw row drilldownsource row, percentile, last updated, eligibility

Source: LiveBench
Raw value: 70.2%
Percentile: 56.9%
Last updated: stale
Eligibility: headline eligible

Derived from the official LiveBench website leaderboard table. Category: IF. Tasks scored: 4.

56.9% percentile inside its fair comparison set

70.2%Raw benchmark value

Language

LB · Chat / text · Objective

It tests whether the model is actually useful in normal conversational turns, not just on narrow correctness tasks.

Rank #25 · Source label: gpt-5.4-xhigh

verified runtimeexact alias

Raw row drilldownsource row, percentile, last updated, eligibility

Source: LiveBench
Raw value: 82.6%
Percentile: 63.1%
Last updated: stale
Eligibility: headline eligible

Derived from the official LiveBench website leaderboard table. Category: Language. Tasks scored: 3.

63.1% percentile inside its fair comparison set

82.6%Raw benchmark value

Paraphrase

LB · Chat / text · Objective

It tests whether the model is actually useful in normal conversational turns, not just on narrow correctness tasks.

Rank #45 · Source label: gpt-5.4-xhigh

verified runtimeexact alias

Raw row drilldownsource row, percentile, last updated, eligibility

Source: LiveBench
Raw value: 63.4%
Percentile: 32.3%
Last updated: stale
Eligibility: headline eligible

Derived from the official LiveBench website leaderboard table. Task: paraphrase. Category: IF.

32.3% percentile inside its fair comparison set

63.4%Raw benchmark value

Simplify

LB · Chat / text · Objective

It tests whether the model is actually useful in normal conversational turns, not just on narrow correctness tasks.

Rank #15 · Source label: gpt-5.4-xhigh

verified runtimeexact alias

Raw row drilldownsource row, percentile, last updated, eligibility

Source: LiveBench
Raw value: 70%
Percentile: 78.5%
Last updated: stale
Eligibility: headline eligible

Derived from the official LiveBench website leaderboard table. Task: simplify. Category: IF.

78.5% percentile inside its fair comparison set

70%Raw benchmark value

Story generation

LB · Chat / text · Objective

It tests whether the model is actually useful in normal conversational turns, not just on narrow correctness tasks.

Rank #5 · Source label: gpt-5.4-xhigh

verified runtimeexact alias

Raw row drilldownsource row, percentile, last updated, eligibility

Source: LiveBench
Raw value: 79%
Percentile: 93.8%
Last updated: stale
Eligibility: headline eligible

Derived from the official LiveBench website leaderboard table. Task: story_generation. Category: IF.

93.8% percentile inside its fair comparison set

79%Raw benchmark value

Summarize

LB · Chat / text · Objective

It tests whether the model is actually useful in normal conversational turns, not just on narrow correctness tasks.

Rank #42 · Source label: gpt-5.4-xhigh

verified runtimeexact alias

Raw row drilldownsource row, percentile, last updated, eligibility

Source: LiveBench
Raw value: 68.4%
Percentile: 36.9%
Last updated: stale
Eligibility: headline eligible

Derived from the official LiveBench website leaderboard table. Task: summarize. Category: IF.

36.9% percentile inside its fair comparison set

68.4%Raw benchmark value

Connections

LB · Chat / text · Objective

It tests whether the model is actually useful in normal conversational turns, not just on narrow correctness tasks.

Rank #3 · Source label: gpt-5.4-xhigh

verified runtimeexact alias

Raw row drilldownsource row, percentile, last updated, eligibility

Source: LiveBench
Raw value: 100%
Percentile: 100%
Last updated: stale
Eligibility: headline eligible

Derived from the official LiveBench website leaderboard table. Task: connections. Category: Language.

100% percentile inside its fair comparison set

100%Raw benchmark value

Plot unscrambling

LB · Chat / text · Objective

It tests whether the model is actually useful in normal conversational turns, not just on narrow correctness tasks.

Rank #26 · Source label: gpt-5.4-xhigh

verified runtimeexact alias

Raw row drilldownsource row, percentile, last updated, eligibility

Source: LiveBench
Raw value: 65.9%
Percentile: 61.5%
Last updated: stale
Eligibility: headline eligible

Derived from the official LiveBench website leaderboard table. Task: plot_unscrambling. Category: Language.

61.5% percentile inside its fair comparison set

65.9%Raw benchmark value

Typos

LB · Chat / text · Objective

It tests whether the model is actually useful in normal conversational turns, not just on narrow correctness tasks.

Rank #28 · Source label: gpt-5.4-xhigh

verified runtimeexact alias

Raw row drilldownsource row, percentile, last updated, eligibility

Source: LiveBench
Raw value: 82%
Percentile: 61.5%
Last updated: stale
Eligibility: headline eligible

Derived from the official LiveBench website leaderboard table. Task: typos. Category: Language.

61.5% percentile inside its fair comparison set

82%Raw benchmark value

Coding31 benchmarks62.7%

Terminal-Bench Hard

AA · Coding · Objective

It tells you whether the model can generate, repair, and reason over code under evaluator pressure rather than marketing examples.

Rank #31 · Source label: GPT-5.4 (Non-reasoning)

verified runtimeexact alias

Raw row drilldownsource row, percentile, last updated, eligibility

Source: Artificial Analysis
Raw value: 37.9%
Percentile: 91%
Last updated: recent
Eligibility: headline eligible

Parsed from Artificial Analysis public leaderboard field `terminalbenchHard`.

91% percentile inside its fair comparison set

37.9%Raw benchmark value

HiL-Bench

SL · Coding · Rubric

It tells you whether the model can generate, repair, and reason over code under evaluator pressure rather than marketing examples.

Rank #15 · Source label: GPT-5.4

verified runtimeexact alias

Raw row drilldownsource row, percentile, last updated, eligibility

Source: Scale Labs
Raw value: 9.7%
Percentile: 12.5%
Last updated: recent
Eligibility: headline eligible

Parsed from the public Scale Labs page for scale-hil-bench. Reported model configuration: GPT-5.4. Collapse policy: highest reported score per canonical model.

12.5% percentile inside its fair comparison set

9.7%Raw benchmark value

Text Arena · Coding

AR · Coding · Human

It tells you whether the model can generate, repair, and reason over code under evaluator pressure rather than marketing examples.

Rank #35 · Source label: gpt-5.4-high

verified runtimeexact alias

Raw row drilldownsource row, percentile, last updated, eligibility

Source: Arena
Raw value: 1,521
Percentile: 90.8%
Last updated: recent
Eligibility: headline eligible

Parsed from Arena leaderboard dataset row `gpt-5.4-high`. Category: coding. Source rank: #35. Votes: 17566. Organization: openai. License: Proprietary.

90.8% percentile inside its fair comparison set

1,521Raw benchmark valueCI 1,515 - 1,527

Text Arena · Coding · No Style Control

AR · Coding · Human

It tells you whether the model can generate, repair, and reason over code under evaluator pressure rather than marketing examples.

Rank #28 · Source label: gpt-5.4-high

verified runtimeexact alias

Raw row drilldownsource row, percentile, last updated, eligibility

Source: Arena
Raw value: 1,497
Percentile: 92.7%
Last updated: recent
Eligibility: headline eligible

Parsed from Arena leaderboard dataset row `gpt-5.4-high`. Category: coding. Source rank: #28. Votes: 17566. Organization: openai. License: Proprietary.

92.7% percentile inside its fair comparison set

1,497Raw benchmark valueCI 1,491 - 1,502

Vibe Code Bench v1.1

VALS-AI · Coding · Objective

It tells you whether the model can generate, repair, and reason over code under evaluator pressure rather than marketing examples.

Rank #35 · Source label: openai/gpt-5.4-2026-03-05

verified runtimeexact alias

Raw row drilldownsource row, percentile, last updated, eligibility

Source: Vals AI
Raw value: 67.4%
Percentile: 63.4%
Last updated: recent
Eligibility: headline eligible

Parsed from Vals AI BenchmarkView overall scores. Vals slug: vibe-code; provider: OpenAI.

63.4% percentile inside its fair comparison set

67.4%Raw benchmark valueCI 57.9% - 76.9%

WebDev Arena

AR · Coding · Human

It tells you whether the model can generate, repair, and reason over code under evaluator pressure rather than marketing examples.

Rank #61 · Source label: gpt-5.4-high (codex-harness)

verified runtimeexact alias

Raw row drilldownsource row, percentile, last updated, eligibility

Source: Arena
Raw value: 1,465
Percentile: 52.8%
Last updated: recent
Eligibility: headline eligible

Parsed from Arena leaderboard dataset row `gpt-5.4-high (codex-harness)`. Category: overall. Source rank: #63. Votes: 1351. Organization: openai. License: Proprietary.

52.8% percentile inside its fair comparison set

1,465Raw benchmark valueCI 1,446 - 1,484

Code Arena · Webdev Html

AR · Coding · Human

It tells you whether the model can generate, repair, and reason over code under evaluator pressure rather than marketing examples.

Rank #58 · Source label: gpt-5.4-medium (codex-harness)

verified runtimeexact alias

Raw row drilldownsource row, percentile, last updated, eligibility

Source: Arena
Raw value: 1,478
Percentile: 54.4%
Last updated: recent
Eligibility: headline eligible

Parsed from Arena leaderboard dataset row `gpt-5.4-medium (codex-harness)`. Category: webdev-html. Source rank: #61. Votes: 132. Organization: openai. License: Proprietary.

54.4% percentile inside its fair comparison set

1,478Raw benchmark valueCI 1,425 - 1,531

Code Arena · Webdev React

AR · Coding · Human

It tells you whether the model can generate, repair, and reason over code under evaluator pressure rather than marketing examples.

Rank #61 · Source label: gpt-5.4-high (codex-harness)

verified runtimeexact alias

Raw row drilldownsource row, percentile, last updated, eligibility

Source: Arena
Raw value: 1,455
Percentile: 46.9%
Last updated: recent
Eligibility: headline eligible

Parsed from Arena leaderboard dataset row `gpt-5.4-high (codex-harness)`. Category: webdev-react. Source rank: #63. Votes: 1235. Organization: openai. License: Proprietary.

46.9% percentile inside its fair comparison set

1,455Raw benchmark valueCI 1,435 - 1,475

Code Migration

VALS-AI · Coding · Objective

It tells you whether the model can generate, repair, and reason over code under evaluator pressure rather than marketing examples.

Rank #33 · Source label: openai/gpt-5.4-2026-03-05

verified runtimeexact alias

Raw row drilldownsource row, percentile, last updated, eligibility

Source: Vals AI
Raw value: 35%
Percentile: 52.9%
Last updated: recent
Eligibility: headline eligible

Parsed from Vals AI BenchmarkView overall scores. Vals slug: code-migration; provider: OpenAI.

52.9% percentile inside its fair comparison set

35%Raw benchmark valueCI 26.9% - 43%

ProgramBench

VALS-AI · Coding · Objective

It tells you whether the model can generate, repair, and reason over code under evaluator pressure rather than marketing examples.

Rank #25 · Source label: openai/gpt-5.4-2026-03-05-high

verified runtimeexact alias

Raw row drilldownsource row, percentile, last updated, eligibility

Source: Vals AI
Raw value: 0.5%
Percentile: 72.4%
Last updated: recent
Eligibility: headline eligible

Parsed from Vals AI BenchmarkView overall scores. Vals slug: programbench; provider: OpenAI.

72.4% percentile inside its fair comparison set

0.5%Raw benchmark valueCI 0% - 1.5%

Code Arena · Image To Webdev

AR · Coding · Human

It tells you whether the model can generate, repair, and reason over code under evaluator pressure rather than marketing examples.

Rank #35 · Source label: gpt-5.4

verified runtimeexact alias

Raw row drilldownsource row, percentile, last updated, eligibility

Source: Arena
Raw value: 1,445
Percentile: 22.7%
Last updated: recent
Eligibility: headline eligible

Parsed from Arena leaderboard dataset row `gpt-5.4`. Category: image_to_webdev. Source rank: #38. Votes: 2433. Organization: openai. License: Proprietary.

22.7% percentile inside its fair comparison set

1,445Raw benchmark valueCI 1,432 - 1,459

SWE-bench Verified

VALS-AI · Coding · Objective

It tells you whether the model can generate, repair, and reason over code under evaluator pressure rather than marketing examples.

Rank #28 · Source label: openai/gpt-5.4-2026-03-05

verified runtimeexact alias

Raw row drilldownsource row, percentile, last updated, eligibility

Source: Vals AI
Raw value: 78.2%
Percentile: 67.1%
Last updated: recent
Eligibility: headline eligible

Parsed from Vals AI BenchmarkView overall scores. Vals slug: swebench; provider: OpenAI.

67.1% percentile inside its fair comparison set

78.2%Raw benchmark valueCI 74.6% - 81.8%

LiveCodeBench

VALS-AI · Coding · Objective

It tells you whether the model can generate, repair, and reason over code under evaluator pressure rather than marketing examples.

Rank #39 · Source label: openai/gpt-5.4-2026-03-05

verified runtimeexact alias

Raw row drilldownsource row, percentile, last updated, eligibility

Source: Vals AI
Raw value: 84.1%
Percentile: 69.6%
Last updated: recent
Eligibility: headline eligible

Parsed from Vals AI BenchmarkView overall scores. Vals slug: lcb; provider: OpenAI.

69.6% percentile inside its fair comparison set

84.1%Raw benchmark valueCI 82.1% - 86.2%

Agentic coding

LB · Coding · Objective

It tells you whether the model can generate, repair, and reason over code under evaluator pressure rather than marketing examples.

Rank #34 · Source label: gpt-5.4-xhigh

verified runtimeexact alias

Raw row drilldownsource row, percentile, last updated, eligibility

Source: LiveBench
Raw value: 53.8%
Percentile: 49.2%
Last updated: stale
Eligibility: headline eligible

Derived from the official LiveBench website leaderboard table. Category: Agentic Coding. Tasks scored: 3.

49.2% percentile inside its fair comparison set

53.8%Raw benchmark value

Coding

LB · Coding · Objective

It tells you whether the model can generate, repair, and reason over code under evaluator pressure rather than marketing examples.

Rank #36 · Source label: gpt-5.4-xhigh

verified runtimeexact alias

Raw row drilldownsource row, percentile, last updated, eligibility

Source: LiveBench
Raw value: 77.5%
Percentile: 46.2%
Last updated: stale
Eligibility: headline eligible

Derived from the official LiveBench website leaderboard table. Category: Coding. Tasks scored: 2.

46.2% percentile inside its fair comparison set

77.5%Raw benchmark value

JavaScript

LB · Coding · Objective

It tells you whether the model can generate, repair, and reason over code under evaluator pressure rather than marketing examples.

Rank #17 · Source label: gpt-5.4-xhigh

verified runtimeexact alias

Raw row drilldownsource row, percentile, last updated, eligibility

Source: LiveBench
Raw value: 68.2%
Percentile: 78.5%
Last updated: stale
Eligibility: headline eligible

Derived from the official LiveBench website leaderboard table. Task: javascript. Category: Agentic Coding.

78.5% percentile inside its fair comparison set

68.2%Raw benchmark value

TypeScript

LB · Coding · Objective

It tells you whether the model can generate, repair, and reason over code under evaluator pressure rather than marketing examples.

Rank #27 · Source label: gpt-5.4-xhigh

verified runtimeexact alias

Raw row drilldownsource row, percentile, last updated, eligibility

Source: LiveBench
Raw value: 43.3%
Percentile: 61.5%
Last updated: stale
Eligibility: headline eligible

Derived from the official LiveBench website leaderboard table. Task: typescript. Category: Agentic Coding.

61.5% percentile inside its fair comparison set

43.3%Raw benchmark value

Python

LB · Coding · Objective

It tells you whether the model can generate, repair, and reason over code under evaluator pressure rather than marketing examples.

Rank #42 · Source label: gpt-5.4-xhigh

verified runtimeexact alias

Raw row drilldownsource row, percentile, last updated, eligibility

Source: LiveBench
Raw value: 50%
Percentile: 46.2%
Last updated: stale
Eligibility: headline eligible

Derived from the official LiveBench website leaderboard table. Task: python. Category: Agentic Coding.

46.2% percentile inside its fair comparison set

50%Raw benchmark value

Coding generation

LB · Coding · Objective

It tells you whether the model can generate, repair, and reason over code under evaluator pressure rather than marketing examples.

Rank #47 · Source label: gpt-5.4-xhigh

verified runtimeexact alias

Raw row drilldownsource row, percentile, last updated, eligibility

Source: LiveBench
Raw value: 74.6%
Percentile: 30.8%
Last updated: stale
Eligibility: headline eligible

Derived from the official LiveBench website leaderboard table. Task: code_generation. Category: Coding.

30.8% percentile inside its fair comparison set

74.6%Raw benchmark value

Coding completion

LB · Coding · Objective

It tells you whether the model can generate, repair, and reason over code under evaluator pressure rather than marketing examples.

Rank #17 · Source label: gpt-5.4-xhigh

verified runtimeexact alias

Raw row drilldownsource row, percentile, last updated, eligibility

Source: LiveBench
Raw value: 80.4%
Percentile: 78.5%
Last updated: stale
Eligibility: headline eligible

Derived from the official LiveBench website leaderboard table. Task: code_completion. Category: Coding.

78.5% percentile inside its fair comparison set

80.4%Raw benchmark value

Debugging

BB · Coding · Rubric

It tells you whether the model can generate, repair, and reason over code under evaluator pressure rather than marketing examples.

Rank #9 · Source label: openai/gpt-5.4

verified runtimeexact alias

Raw row drilldownsource row, percentile, last updated, eligibility

Source: BridgeBench
Raw value: 85.6%
Percentile: 66.7%
Last updated: archived
Eligibility: headline eligible

Parsed from the BridgeBench page for bridgebench-debugging.

66.7% percentile inside its fair comparison set

85.6%Raw benchmark value

Refactoring

BB · Coding · Rubric

It tells you whether the model can generate, repair, and reason over code under evaluator pressure rather than marketing examples.

Rank #12 · Source label: openai/gpt-5.4

verified runtimeexact alias

Raw row drilldownsource row, percentile, last updated, eligibility

Source: BridgeBench
Raw value: 63.4%
Percentile: 47.6%
Last updated: archived
Eligibility: headline eligible

Parsed from the BridgeBench page for bridgebench-refactoring.

47.6% percentile inside its fair comparison set

63.4%Raw benchmark value

Security

BB · Coding · Rubric

It tells you whether the model can generate, repair, and reason over code under evaluator pressure rather than marketing examples.

Rank #4 · Source label: openai/gpt-5.4

verified runtimeexact alias

Raw row drilldownsource row, percentile, last updated, eligibility

Source: BridgeBench
Raw value: 84.4%
Percentile: 88%
Last updated: archived
Eligibility: headline eligible

Parsed from the BridgeBench page for bridgebench-security.

88% percentile inside its fair comparison set

84.4%Raw benchmark value

UI bench

BB · Coding · Rubric

It tells you whether the model can generate, repair, and reason over code under evaluator pressure rather than marketing examples.

Rank #11 · Source label: openai/gpt-5.4

verified runtimeexact alias

Raw row drilldownsource row, percentile, last updated, eligibility

Source: BridgeBench
Raw value: 65.7%
Percentile: 54.5%
Last updated: archived
Eligibility: headline eligible

Parsed from the BridgeBench page for bridgebench-ui.

54.5% percentile inside its fair comparison set

65.7%Raw benchmark value

Speed throughput

BB · Coding · Speed / cost

It tells you whether the model can generate, repair, and reason over code under evaluator pressure rather than marketing examples.

Rank #18 · Source label: openai/gpt-5.4

verified runtimeexact alias

Raw row drilldownsource row, percentile, last updated, eligibility

Source: BridgeBench
Raw value: 88 t/s
Percentile: 32%
Last updated: archived
Eligibility: headline eligible

Parsed from the BridgeBench page for bridgebench-throughput.

32% percentile inside its fair comparison set

88 t/sRaw benchmark value

Speed TTFT

BB · Coding · Speed / cost

It tells you whether the model can generate, repair, and reason over code under evaluator pressure rather than marketing examples.

Rank #2 · Source label: openai/gpt-5.4

verified runtimeexact alias

Raw row drilldownsource row, percentile, last updated, eligibility

Source: BridgeBench
Raw value: 397.00ms
Percentile: 96%
Last updated: archived
Eligibility: headline eligible

Parsed from the BridgeBench page for bridgebench-ttft.

96% percentile inside its fair comparison set

397.00msRaw benchmark value

Speed cost

BB · Coding · Speed / cost

It tells you whether the model can generate, repair, and reason over code under evaluator pressure rather than marketing examples.

Rank #21 · Source label: openai/gpt-5.4

verified runtimeexact alias

Raw row drilldownsource row, percentile, last updated, eligibility

Source: BridgeBench
Raw value: $0.3 $
Percentile: 20%
Last updated: archived
Eligibility: headline eligible

Parsed from the BridgeBench page for bridgebench-speed-cost.

20% percentile inside its fair comparison set

$0.3 $Raw benchmark value

Terminal-Bench 2.0

OFF · Coding · Objective

It tells you whether the model can generate, repair, and reason over code under evaluator pressure rather than marketing examples.

Rank #2

Official company resultmanual verifiedmanual verified

Raw row drilldownsource row, percentile, last updated, eligibility

Source: Provider official evals
Raw value: 75.1%
Percentile: 83.3%
Last updated: archived
Eligibility: headline eligible

Verified against OpenAI's GPT-5.4 launch page dated March 5, 2026. Uses the published Terminal-Bench 2.0 figure. Checked 2026-04-29. Verification: manual_public_page_verification.

83.3% percentile inside its fair comparison set

75.1%Raw benchmark value

Overall quality

BB · Coding · Combined

It tells you whether the model can generate, repair, and reason over code under evaluator pressure rather than marketing examples.

Rank #8 · Source label: GPT-5.4

verified runtimeexact alias

Raw row drilldownsource row, percentile, last updated, eligibility

Source: BridgeBench
Raw value: 86.3%
Percentile: 68.2%
Last updated: archived
Eligibility: headline eligible

Parsed from the BridgeBench page for bridgebench-overall-quality.

68.2% percentile inside its fair comparison set

86.3%Raw benchmark value

Overall vibe

BB · Coding · Combined

It tells you whether the model can generate, repair, and reason over code under evaluator pressure rather than marketing examples.

Rank #9 · Source label: GPT-5.4

verified runtimeexact alias

Raw row drilldownsource row, percentile, last updated, eligibility

Source: BridgeBench
Raw value: 79.6%
Percentile: 63.6%
Last updated: archived
Eligibility: headline eligible

Parsed from the BridgeBench page for bridgebench-overall-vibe.

63.6% percentile inside its fair comparison set

79.6%Raw benchmark value

Overall coverage

BB · Coding · Combined

It tells you whether the model can generate, repair, and reason over code under evaluator pressure rather than marketing examples.

Rank #7 · Source label: GPT-5.4

verified runtimeexact alias

Raw row drilldownsource row, percentile, last updated, eligibility

Source: BridgeBench
Raw value: 100%
Percentile: 100%
Last updated: archived
Eligibility: headline eligible

Parsed from the BridgeBench page for bridgebench-overall-coverage.

100% percentile inside its fair comparison set

100%Raw benchmark value

Reasoning / math / science21 benchmarks72.7%

Humanity's Last Exam

AA · Reasoning / math / science · Objective

It is one of the cleaner reads on deliberate reasoning strength rather than style or popularity.

Rank #210 · Source label: GPT-5.4 (Non-reasoning)

verified runtimeexact alias

Raw row drilldownsource row, percentile, last updated, eligibility

Source: Artificial Analysis
Raw value: 11.3%
Percentile: 60.5%
Last updated: recent
Eligibility: headline eligible

Parsed from Artificial Analysis public leaderboard field `hle`.

60.5% percentile inside its fair comparison set

11.3%Raw benchmark value

GPQA

AA · Reasoning / math / science · Objective

It is one of the cleaner reads on deliberate reasoning strength rather than style or popularity.

Rank #183 · Source label: GPT-5.4 (Non-reasoning)

verified runtimeexact alias

Raw row drilldownsource row, percentile, last updated, eligibility

Source: Artificial Analysis
Raw value: 74.8%
Percentile: 63.1%
Last updated: recent
Eligibility: headline eligible

Parsed from Artificial Analysis public leaderboard field `gpqa`.

63.1% percentile inside its fair comparison set

74.8%Raw benchmark value

CritPt

AA · Reasoning / math / science · Objective

It is one of the cleaner reads on deliberate reasoning strength rather than style or popularity.

Rank #171 · Source label: GPT-5.4 (Non-reasoning)

verified runtimeexact alias

Raw row drilldownsource row, percentile, last updated, eligibility

Source: Artificial Analysis
Raw value: 0.6%
Percentile: 62.8%
Last updated: recent
Eligibility: headline eligible

Parsed from Artificial Analysis public leaderboard field `critpt`.

62.8% percentile inside its fair comparison set

0.6%Raw benchmark value

TutorBench

SL · Reasoning / math / science · Rubric

It is one of the cleaner reads on deliberate reasoning strength rather than style or popularity.

Rank #2 · Source label: gpt-5.4-pro-2026-03-05

verified runtimeexact alias

Raw row drilldownsource row, percentile, last updated, eligibility

Source: Scale Labs
Raw value: 56.6%
Percentile: 96.3%
Last updated: recent
Eligibility: headline eligible

Parsed from the public Scale Labs page for scale-tutorbench. Reported model configuration: gpt-5.4-pro-2026-03-05. Collapse policy: highest reported score per canonical model.

96.3% percentile inside its fair comparison set

56.6%Raw benchmark value

MultiNRC

SL · Reasoning / math / science · Rubric

It is one of the cleaner reads on deliberate reasoning strength rather than style or popularity.

Rank #6 · Source label: gpt-5.4-pro-2026-03-05

verified runtimeexact alias

Raw row drilldownsource row, percentile, last updated, eligibility

Source: Scale Labs
Raw value: 62.3%
Percentile: 88.6%
Last updated: recent
Eligibility: headline eligible

Parsed from the public Scale Labs page for scale-multinrc. Reported model configuration: gpt-5.4-pro-2026-03-05. Collapse policy: highest reported score per canonical model.

88.6% percentile inside its fair comparison set

62.3%Raw benchmark value

Text Arena · Math

AR · Reasoning / math / science · Human

It is one of the cleaner reads on deliberate reasoning strength rather than style or popularity.

Rank #24 · Source label: gpt-5.4-high

verified runtimeexact alias

Raw row drilldownsource row, percentile, last updated, eligibility

Source: Arena
Raw value: 1,492
Percentile: 93.6%
Last updated: recent
Eligibility: headline eligible

Parsed from Arena leaderboard dataset row `gpt-5.4-high`. Category: math. Source rank: #25. Votes: 3379. Organization: openai. License: Proprietary.

93.6% percentile inside its fair comparison set

1,492Raw benchmark valueCI 1,481 - 1,502

Text Arena · Math · No Style Control

AR · Reasoning / math / science · Human

It is one of the cleaner reads on deliberate reasoning strength rather than style or popularity.

Rank #22 · Source label: gpt-5.4-high

verified runtimeexact alias

Raw row drilldownsource row, percentile, last updated, eligibility

Source: Arena
Raw value: 1,488
Percentile: 94.2%
Last updated: recent
Eligibility: headline eligible

Parsed from Arena leaderboard dataset row `gpt-5.4-high`. Category: math. Source rank: #22. Votes: 3379. Organization: openai. License: Proprietary.

94.2% percentile inside its fair comparison set

1,488Raw benchmark valueCI 1,477 - 1,498

GPQA Diamond

VALS-AI · Reasoning / math / science · Objective

It is one of the cleaner reads on deliberate reasoning strength rather than style or popularity.

Rank #21 · Source label: openai/gpt-5.4-2026-03-05

verified runtimeexact alias

Raw row drilldownsource row, percentile, last updated, eligibility

Source: Vals AI
Raw value: 91.7%
Percentile: 86.3%
Last updated: recent
Eligibility: headline eligible

Parsed from Vals AI BenchmarkView overall scores. Vals slug: gpqa; provider: OpenAI.

86.3% percentile inside its fair comparison set

91.7%Raw benchmark valueCI 87.9% - 95.4%

MMLU Pro

VALS-AI · Reasoning / math / science · Objective

It is one of the cleaner reads on deliberate reasoning strength rather than style or popularity.

Rank #27 · Source label: openai/gpt-5.4-2026-03-05

verified runtimeexact alias

Raw row drilldownsource row, percentile, last updated, eligibility

Source: Vals AI
Raw value: 87.5%
Percentile: 79.2%
Last updated: recent
Eligibility: headline eligible

Parsed from Vals AI BenchmarkView overall scores. Vals slug: mmlu_pro; provider: OpenAI.

79.2% percentile inside its fair comparison set

87.5%Raw benchmark valueCI 86.7% - 88.3%

Mathematics

LB · Reasoning / math / science · Objective

It is one of the cleaner reads on deliberate reasoning strength rather than style or popularity.

Rank #19 · Source label: gpt-5.4-xhigh

verified runtimeexact alias

Raw row drilldownsource row, percentile, last updated, eligibility

Source: LiveBench
Raw value: 94.1%
Percentile: 72.3%
Last updated: stale
Eligibility: headline eligible

Derived from the official LiveBench website leaderboard table. Category: Mathematics. Tasks scored: 4.

72.3% percentile inside its fair comparison set

94.1%Raw benchmark value

Reasoning

LB · Reasoning / math / science · Objective

It is one of the cleaner reads on deliberate reasoning strength rather than style or popularity.

Rank #24 · Source label: gpt-5.4-xhigh

verified runtimeexact alias

Raw row drilldownsource row, percentile, last updated, eligibility

Source: LiveBench
Raw value: 88.1%
Percentile: 64.6%
Last updated: stale
Eligibility: headline eligible

Derived from the official LiveBench website leaderboard table. Category: Reasoning. Tasks scored: 4.

64.6% percentile inside its fair comparison set

88.1%Raw benchmark value

AMPS Hard

LB · Reasoning / math / science · Objective

It is one of the cleaner reads on deliberate reasoning strength rather than style or popularity.

Rank #21 · Source label: gpt-5.4-xhigh

verified runtimeexact alias

Raw row drilldownsource row, percentile, last updated, eligibility

Source: LiveBench
Raw value: 98%
Percentile: 83.1%
Last updated: stale
Eligibility: headline eligible

Derived from the official LiveBench website leaderboard table. Task: AMPS_Hard. Category: Mathematics.

83.1% percentile inside its fair comparison set

98%Raw benchmark value

Integrals with game

LB · Reasoning / math / science · Objective

It is one of the cleaner reads on deliberate reasoning strength rather than style or popularity.

Rank #18 · Source label: gpt-5.4-xhigh

verified runtimeexact alias

Raw row drilldownsource row, percentile, last updated, eligibility

Source: LiveBench
Raw value: 93%
Percentile: 73.8%
Last updated: stale
Eligibility: headline eligible

Derived from the official LiveBench website leaderboard table. Task: integrals_with_game. Category: Mathematics.

73.8% percentile inside its fair comparison set

93%Raw benchmark value

Math competition

LB · Reasoning / math / science · Objective

It is one of the cleaner reads on deliberate reasoning strength rather than style or popularity.

Rank #51 · Source label: gpt-5.4-xhigh

verified runtimeexact alias

Raw row drilldownsource row, percentile, last updated, eligibility

Source: LiveBench
Raw value: 94.1%
Percentile: 26.2%
Last updated: stale
Eligibility: headline eligible

Derived from the official LiveBench website leaderboard table. Task: math_comp. Category: Mathematics.

26.2% percentile inside its fair comparison set

94.1%Raw benchmark value

Olympiad

LB · Reasoning / math / science · Objective

It is one of the cleaner reads on deliberate reasoning strength rather than style or popularity.

Rank #23 · Source label: gpt-5.4-xhigh

verified runtimeexact alias

Raw row drilldownsource row, percentile, last updated, eligibility

Source: LiveBench
Raw value: 91.5%
Percentile: 66.2%
Last updated: stale
Eligibility: headline eligible

Derived from the official LiveBench website leaderboard table. Task: olympiad. Category: Mathematics.

66.2% percentile inside its fair comparison set

91.5%Raw benchmark value

Theory of mind

LB · Reasoning / math / science · Objective

It is one of the cleaner reads on deliberate reasoning strength rather than style or popularity.

Rank #1 · Source label: gpt-5.4-xhigh

verified runtimeexact alias

Raw row drilldownsource row, percentile, last updated, eligibility

Source: LiveBench
Raw value: 88.5%
Percentile: 100%
Last updated: stale
Eligibility: headline eligible

Derived from the official LiveBench website leaderboard table. Task: theory_of_mind. Category: Reasoning.

100% percentile inside its fair comparison set

88.5%Raw benchmark value

Zebra puzzle

LB · Reasoning / math / science · Objective

It is one of the cleaner reads on deliberate reasoning strength rather than style or popularity.

Rank #25 · Source label: gpt-5.4-xhigh

verified runtimeexact alias

Raw row drilldownsource row, percentile, last updated, eligibility

Source: LiveBench
Raw value: 98%
Percentile: 64.6%
Last updated: stale
Eligibility: headline eligible

Derived from the official LiveBench website leaderboard table. Task: zebra_puzzle. Category: Reasoning.

64.6% percentile inside its fair comparison set

98%Raw benchmark value

Spatial

LB · Reasoning / math / science · Objective

It is one of the cleaner reads on deliberate reasoning strength rather than style or popularity.

Rank #21 · Source label: gpt-5.4-xhigh

verified runtimeexact alias

Raw row drilldownsource row, percentile, last updated, eligibility

Source: LiveBench
Raw value: 98%
Percentile: 72.3%
Last updated: stale
Eligibility: headline eligible

Derived from the official LiveBench website leaderboard table. Task: spatial. Category: Reasoning.

72.3% percentile inside its fair comparison set

98%Raw benchmark value

Logic with navigation

LB · Reasoning / math / science · Objective

It is one of the cleaner reads on deliberate reasoning strength rather than style or popularity.

Rank #48 · Source label: gpt-5.4-xhigh

verified runtimeexact alias

Raw row drilldownsource row, percentile, last updated, eligibility

Source: LiveBench
Raw value: 68%
Percentile: 30.8%
Last updated: stale
Eligibility: headline eligible

Derived from the official LiveBench website leaderboard table. Task: logic_with_navigation. Category: Reasoning.

30.8% percentile inside its fair comparison set

68%Raw benchmark value

Reasoning

BB · Reasoning / math / science · Rubric

It is one of the cleaner reads on deliberate reasoning strength rather than style or popularity.

Rank #3 · Source label: openai/gpt-5.4

verified runtimeexact alias

Raw row drilldownsource row, percentile, last updated, eligibility

Source: BridgeBench
Raw value: 40.6%
Percentile: 90.5%
Last updated: archived
Eligibility: headline eligible

Parsed from the BridgeBench page for bridgebench-reasoning.

90.5% percentile inside its fair comparison set

40.6%Raw benchmark value

Humanity's Last Exam

OFF · Reasoning / math / science · Objective

It is one of the cleaner reads on deliberate reasoning strength rather than style or popularity.

Rank #5

Official company resultmanual verifiedmanual verified

Raw row drilldownsource row, percentile, last updated, eligibility

Source: Provider official evals
Raw value: 39.8%
Percentile: 42.9%
Last updated: archived
Eligibility: headline eligible

Verified against OpenAI's GPT-5.4 launch page dated March 5, 2026. Uses the published no-tools Humanity's Last Exam figure. Checked 2026-04-29. Verification: manual_public_page_verification.

42.9% percentile inside its fair comparison set

39.8%Raw benchmark value

Professional reasoning35 benchmarks66.6%

APEX-Agents-AA

AA · Professional reasoning · Objective

Long-horizon agentic task completion.

Rank #9 · Source label: GPT-5.4 (Xhigh)

verified runtimeexact alias

Raw row drilldownsource row, percentile, last updated, eligibility

Source: Artificial Analysis
Raw value: 33.3%
Percentile: 75%
Last updated: recent
Eligibility: headline eligible

Parsed from Artificial Analysis public leaderboard field `apexAgents`.

75% percentile inside its fair comparison set

33.3%Raw benchmark value

PRBench Legal

SL · Professional reasoning · Rubric

Applied legal reasoning on professional-domain tasks.

Rank #19 · Source label: gpt-5.4 (High)

verified runtimeexact alias

Raw row drilldownsource row, percentile, last updated, eligibility

Source: Scale Labs
Raw value: 44.4%
Percentile: 56.1%
Last updated: recent
Eligibility: headline eligible

Parsed from the public Scale Labs page for scale-prbench-legal. Reported model configuration: gpt-5.4 (High). Collapse policy: highest reported score per canonical model.

56.1% percentile inside its fair comparison set

44.4%Raw benchmark value

Text Arena · Expert

AR · Professional reasoning · Human

Observed user preference in Arena's Text Arena expert leaderboard.

Rank #23 · Source label: gpt-5.4-high

verified runtimeexact alias

Raw row drilldownsource row, percentile, last updated, eligibility

Source: Arena
Raw value: 1,515
Percentile: 93.3%
Last updated: recent
Eligibility: headline eligible

Parsed from Arena leaderboard dataset row `gpt-5.4-high`. Category: expert. Source rank: #23. Votes: 6333. Organization: openai. License: Proprietary.

93.3% percentile inside its fair comparison set

1,515Raw benchmark valueCI 1,507 - 1,524

Text Arena · Industry Business And Management And Financial Operations

AR · Professional reasoning · Human

Observed user preference in Arena's Text Arena industry_business_and_management_and_financial_operations leaderboard.

Rank #24 · Source label: gpt-5.4-high

verified runtimeexact alias

Raw row drilldownsource row, percentile, last updated, eligibility

Source: Arena
Raw value: 1,482
Percentile: 93.8%
Last updated: recent
Eligibility: headline eligible

Parsed from Arena leaderboard dataset row `gpt-5.4-high`. Category: industry_business_and_management_and_financial_operations. Source rank: #25. Votes: 13075. Organization: openai. License: Proprietary.

93.8% percentile inside its fair comparison set

1,482Raw benchmark valueCI 1,475 - 1,488

Text Arena · Industry Entertainment And Sports And Media

AR · Professional reasoning · Human

Observed user preference in Arena's Text Arena industry_entertainment_and_sports_and_media leaderboard.

Rank #47 · Source label: gpt-5.4-high

verified runtimeexact alias

Raw row drilldownsource row, percentile, last updated, eligibility

Source: Arena
Raw value: 1,441
Percentile: 87.7%
Last updated: recent
Eligibility: headline eligible

Parsed from Arena leaderboard dataset row `gpt-5.4-high`. Category: industry_entertainment_and_sports_and_media. Source rank: #50. Votes: 14124. Organization: openai. License: Proprietary.

87.7% percentile inside its fair comparison set

1,441Raw benchmark valueCI 1,435 - 1,448

Text Arena · Industry Legal And Government

AR · Professional reasoning · Human

Observed user preference in Arena's Text Arena industry_legal_and_government leaderboard.

Rank #26 · Source label: gpt-5.4-high

verified runtimeexact alias

Raw row drilldownsource row, percentile, last updated, eligibility

Source: Arena
Raw value: 1,490
Percentile: 92.8%
Last updated: recent
Eligibility: headline eligible

Parsed from Arena leaderboard dataset row `gpt-5.4-high`. Category: industry_legal_and_government. Source rank: #26. Votes: 5228. Organization: openai. License: Proprietary.

92.8% percentile inside its fair comparison set

1,490Raw benchmark valueCI 1,481 - 1,499

Text Arena · Industry Life And Physical And Social Science

AR · Professional reasoning · Human

Observed user preference in Arena's Text Arena industry_life_and_physical_and_social_science leaderboard.

Rank #42 · Source label: gpt-5.4-high

verified runtimeexact alias

Raw row drilldownsource row, percentile, last updated, eligibility

Source: Arena
Raw value: 1,488
Percentile: 89%
Last updated: recent
Eligibility: headline eligible

Parsed from Arena leaderboard dataset row `gpt-5.4-high`. Category: industry_life_and_physical_and_social_science. Source rank: #44. Votes: 10640. Organization: openai. License: Proprietary.

89% percentile inside its fair comparison set

1,488Raw benchmark valueCI 1,481 - 1,495

Text Arena · Industry Mathematical

AR · Professional reasoning · Human

Observed user preference in Arena's Text Arena industry_mathematical leaderboard.

Rank #19 · Source label: gpt-5.4-high

verified runtimeexact alias

Raw row drilldownsource row, percentile, last updated, eligibility

Source: Arena
Raw value: 1,499
Percentile: 94.9%
Last updated: recent
Eligibility: headline eligible

Parsed from Arena leaderboard dataset row `gpt-5.4-high`. Category: industry_mathematical. Source rank: #19. Votes: 3543. Organization: openai. License: Proprietary.

94.9% percentile inside its fair comparison set

1,499Raw benchmark valueCI 1,488 - 1,510

Text Arena · Industry Medicine And Healthcare

AR · Professional reasoning · Human

Observed user preference in Arena's Text Arena industry_medicine_and_healthcare leaderboard.

Rank #60 · Source label: gpt-5.4-high

verified runtimeexact alias

Raw row drilldownsource row, percentile, last updated, eligibility

Source: Arena
Raw value: 1,475
Percentile: 82.9%
Last updated: recent
Eligibility: headline eligible

Parsed from Arena leaderboard dataset row `gpt-5.4-high`. Category: industry_medicine_and_healthcare. Source rank: #66. Votes: 4838. Organization: openai. License: Proprietary.

82.9% percentile inside its fair comparison set

1,475Raw benchmark valueCI 1,465 - 1,485

Text Arena · Industry Software And It Services

AR · Professional reasoning · Human

Observed user preference in Arena's Text Arena industry_software_and_it_services leaderboard.

Rank #41 · Source label: gpt-5.4-high

verified runtimeexact alias

Raw row drilldownsource row, percentile, last updated, eligibility

Source: Arena
Raw value: 1,507
Percentile: 89.4%
Last updated: recent
Eligibility: headline eligible

Parsed from Arena leaderboard dataset row `gpt-5.4-high`. Category: industry_software_and_it_services. Source rank: #41. Votes: 25424. Organization: openai. License: Proprietary.

89.4% percentile inside its fair comparison set

1,507Raw benchmark valueCI 1,502 - 1,512

Text Arena · Industry Writing And Literature And Language

AR · Professional reasoning · Human

Observed user preference in Arena's Text Arena industry_writing_and_literature_and_language leaderboard.

Rank #30 · Source label: gpt-5.4-high

verified runtimeexact alias

Raw row drilldownsource row, percentile, last updated, eligibility

Source: Arena
Raw value: 1,465
Percentile: 92.3%
Last updated: recent
Eligibility: headline eligible

Parsed from Arena leaderboard dataset row `gpt-5.4-high`. Category: industry_writing_and_literature_and_language. Source rank: #32. Votes: 16324. Organization: openai. License: Proprietary.

92.3% percentile inside its fair comparison set

1,465Raw benchmark valueCI 1,459 - 1,471

Text Arena · Expert · No Style Control

AR · Professional reasoning · Human

Observed user preference in Arena's Text Arena expert leaderboard.

Rank #22 · Source label: gpt-5.4-high

verified runtimeexact alias

Raw row drilldownsource row, percentile, last updated, eligibility

Source: Arena
Raw value: 1,507
Percentile: 93.6%
Last updated: recent
Eligibility: headline eligible

Parsed from Arena leaderboard dataset row `gpt-5.4-high`. Category: expert. Source rank: #22. Votes: 6333. Organization: openai. License: Proprietary.

93.6% percentile inside its fair comparison set

1,507Raw benchmark valueCI 1,499 - 1,515

Text Arena · Industry Business And Management And Financial Operations · No Style Control

AR · Professional reasoning · Human

Observed user preference in Arena's Text Arena industry_business_and_management_and_financial_operations leaderboard.

Rank #17 · Source label: gpt-5.4-high

verified runtimeexact alias

Raw row drilldownsource row, percentile, last updated, eligibility

Source: Arena
Raw value: 1,475
Percentile: 95.7%
Last updated: recent
Eligibility: headline eligible

Parsed from Arena leaderboard dataset row `gpt-5.4-high`. Category: industry_business_and_management_and_financial_operations. Source rank: #17. Votes: 13075. Organization: openai. License: Proprietary.

95.7% percentile inside its fair comparison set

1,475Raw benchmark valueCI 1,468 - 1,481

Text Arena · Industry Entertainment And Sports And Media · No Style Control

AR · Professional reasoning · Human

Observed user preference in Arena's Text Arena industry_entertainment_and_sports_and_media leaderboard.

Rank #38 · Source label: gpt-5.4-high

verified runtimeexact alias

Raw row drilldownsource row, percentile, last updated, eligibility

Source: Arena
Raw value: 1,438
Percentile: 90.1%
Last updated: recent
Eligibility: headline eligible

Parsed from Arena leaderboard dataset row `gpt-5.4-high`. Category: industry_entertainment_and_sports_and_media. Source rank: #40. Votes: 14124. Organization: openai. License: Proprietary.

90.1% percentile inside its fair comparison set

1,438Raw benchmark valueCI 1,432 - 1,444

Text Arena · Industry Legal And Government · No Style Control

AR · Professional reasoning · Human

Observed user preference in Arena's Text Arena industry_legal_and_government leaderboard.

Rank #21 · Source label: gpt-5.4-high

verified runtimeexact alias

Raw row drilldownsource row, percentile, last updated, eligibility

Source: Arena
Raw value: 1,486
Percentile: 94.3%
Last updated: recent
Eligibility: headline eligible

Parsed from Arena leaderboard dataset row `gpt-5.4-high`. Category: industry_legal_and_government. Source rank: #22. Votes: 5228. Organization: openai. License: Proprietary.

94.3% percentile inside its fair comparison set

1,486Raw benchmark valueCI 1,477 - 1,495

Text Arena · Industry Life And Physical And Social Science · No Style Control

AR · Professional reasoning · Human

Observed user preference in Arena's Text Arena industry_life_and_physical_and_social_science leaderboard.

Rank #32 · Source label: gpt-5.4-high

verified runtimeexact alias

Raw row drilldownsource row, percentile, last updated, eligibility

Source: Arena
Raw value: 1,478
Percentile: 91.7%
Last updated: recent
Eligibility: headline eligible

Parsed from Arena leaderboard dataset row `gpt-5.4-high`. Category: industry_life_and_physical_and_social_science. Source rank: #34. Votes: 10640. Organization: openai. License: Proprietary.

91.7% percentile inside its fair comparison set

1,478Raw benchmark valueCI 1,472 - 1,485

Text Arena · Industry Mathematical · No Style Control

AR · Professional reasoning · Human

Observed user preference in Arena's Text Arena industry_mathematical leaderboard.

Rank #16 · Source label: gpt-5.4-high

verified runtimeexact alias

Raw row drilldownsource row, percentile, last updated, eligibility

Source: Arena
Raw value: 1,497
Percentile: 95.8%
Last updated: recent
Eligibility: headline eligible

Parsed from Arena leaderboard dataset row `gpt-5.4-high`. Category: industry_mathematical. Source rank: #16. Votes: 3543. Organization: openai. License: Proprietary.

95.8% percentile inside its fair comparison set

1,497Raw benchmark valueCI 1,486 - 1,507

Text Arena · Industry Medicine And Healthcare · No Style Control

AR · Professional reasoning · Human

Observed user preference in Arena's Text Arena industry_medicine_and_healthcare leaderboard.

Rank #52 · Source label: gpt-5.4-high

verified runtimeexact alias

Raw row drilldownsource row, percentile, last updated, eligibility

Source: Arena
Raw value: 1,459
Percentile: 85.3%
Last updated: recent
Eligibility: headline eligible

Parsed from Arena leaderboard dataset row `gpt-5.4-high`. Category: industry_medicine_and_healthcare. Source rank: #53. Votes: 4838. Organization: openai. License: Proprietary.

85.3% percentile inside its fair comparison set

1,459Raw benchmark valueCI 1,449 - 1,468

Text Arena · Industry Software And It Services · No Style Control

AR · Professional reasoning · Human

Observed user preference in Arena's Text Arena industry_software_and_it_services leaderboard.

Rank #30 · Source label: gpt-5.4-high

verified runtimeexact alias

Raw row drilldownsource row, percentile, last updated, eligibility

Source: Arena
Raw value: 1,491
Percentile: 92.3%
Last updated: recent
Eligibility: headline eligible

Parsed from Arena leaderboard dataset row `gpt-5.4-high`. Category: industry_software_and_it_services. Source rank: #30. Votes: 25424. Organization: openai. License: Proprietary.

92.3% percentile inside its fair comparison set

1,491Raw benchmark valueCI 1,486 - 1,496

Text Arena · Industry Writing And Literature And Language · No Style Control

AR · Professional reasoning · Human

Observed user preference in Arena's Text Arena industry_writing_and_literature_and_language leaderboard.

Rank #31 · Source label: gpt-5.4-high

verified runtimeexact alias

Raw row drilldownsource row, percentile, last updated, eligibility

Source: Arena
Raw value: 1,458
Percentile: 92%
Last updated: recent
Eligibility: headline eligible

Parsed from Arena leaderboard dataset row `gpt-5.4-high`. Category: industry_writing_and_literature_and_language. Source rank: #33. Votes: 16324. Organization: openai. License: Proprietary.

92% percentile inside its fair comparison set

1,458Raw benchmark valueCI 1,453 - 1,464

Harvey's Legal Agent Benchmark

VALS-AI · Professional reasoning · Objective

Completing legal work with documents, spreadsheets, presentations, and file-system tools.

Rank #68 · Source label: openai/gpt-5.4-2026-03-05

verified runtimeexact alias

Raw row drilldownsource row, percentile, last updated, eligibility

Source: Vals AI
Raw value: 0%
Percentile: 20.3%
Last updated: recent
Eligibility: headline eligible

Parsed from Vals AI BenchmarkView overall scores. Vals slug: hlab; provider: OpenAI.

20.3% percentile inside its fair comparison set

0%Raw benchmark valueCI 0% - 0%

LegalBench

VALS-AI · Professional reasoning · Objective

Academic legal reasoning tasks.

Rank #15 · Source label: openai/gpt-5.4-2026-03-05

verified runtimeexact alias

Raw row drilldownsource row, percentile, last updated, eligibility

Source: Vals AI
Raw value: 86%
Percentile: 89.6%
Last updated: recent
Eligibility: headline eligible

Parsed from Vals AI BenchmarkView overall scores. Vals slug: legal_bench; provider: OpenAI.

89.6% percentile inside its fair comparison set

86%Raw benchmark valueCI 85.2% - 86.9%

MedCode

VALS-AI · Professional reasoning · Objective

Medical billing support and coding tasks.

Rank #54 · Source label: openai/gpt-5.4-2026-03-05

verified runtimeexact alias

Raw row drilldownsource row, percentile, last updated, eligibility

Source: Vals AI
Raw value: 41.3%
Percentile: 44.2%
Last updated: recent
Eligibility: headline eligible

Parsed from Vals AI BenchmarkView overall scores. Vals slug: medcode; provider: OpenAI.

44.2% percentile inside its fair comparison set

41.3%Raw benchmark valueCI 37.1% - 45.5%

MedScribe

VALS-AI · Professional reasoning · Objective

Administrative documentation support for doctors.

Rank #62 · Source label: openai/gpt-5.4-2026-03-05

verified runtimeexact alias

Raw row drilldownsource row, percentile, last updated, eligibility

Source: Vals AI
Raw value: 77.5%
Percentile: 36.5%
Last updated: recent
Eligibility: headline eligible

Parsed from Vals AI BenchmarkView overall scores. Vals slug: medscribe; provider: OpenAI.

36.5% percentile inside its fair comparison set

77.5%Raw benchmark valueCI 71% - 84%

SAGE

VALS-AI · Professional reasoning · Objective

Student Assessment with Generative Evaluation.

Rank #48 · Source label: openai/gpt-5.4-2026-03-05

verified runtimeexact alias

Raw row drilldownsource row, percentile, last updated, eligibility

Source: Vals AI
Raw value: 43.3%
Percentile: 42%
Last updated: recent
Eligibility: headline eligible

Parsed from Vals AI BenchmarkView overall scores. Vals slug: sage; provider: OpenAI.

42% percentile inside its fair comparison set

43.3%Raw benchmark valueCI 37.2% - 49.4%

SkillsBench

VALS-AI · Professional reasoning · Objective

Applied professional skills tasks.

Rank #18 · Source label: openai/gpt-5.4-2026-03-05

verified runtimeexact alias

Raw row drilldownsource row, percentile, last updated, eligibility

Source: Vals AI
Raw value: 51.7%
Percentile: 48.5%
Last updated: recent
Eligibility: headline eligible

Parsed from Vals AI BenchmarkView overall scores. Vals slug: skillsbench; provider: OpenAI.

48.5% percentile inside its fair comparison set

51.7%Raw benchmark valueCI 42.5% - 61%

TaxEval v2

VALS-AI · Professional reasoning · Objective

Answer quality on tax questions and responses.

Rank #44 · Source label: openai/gpt-5.4-2026-03-05

verified runtimeexact alias

Raw row drilldownsource row, percentile, last updated, eligibility

Source: Vals AI
Raw value: 74%
Percentile: 67.4%
Last updated: recent
Eligibility: headline eligible

Parsed from Vals AI BenchmarkView overall scores. Vals slug: tax_eval_v2; provider: OpenAI.

67.4% percentile inside its fair comparison set

74%Raw benchmark valueCI 72.3% - 75.7%

Data analysis

LB · Professional reasoning · Objective

Structured data manipulation and table reasoning accuracy.

Rank #15 · Source label: gpt-5.4-xhigh

verified runtimeexact alias

Raw row drilldownsource row, percentile, last updated, eligibility

Source: LiveBench
Raw value: 79.3%
Percentile: 78.5%
Last updated: stale
Eligibility: headline eligible

Derived from the official LiveBench website leaderboard table. Category: Data Analysis. Tasks scored: 3.

78.5% percentile inside its fair comparison set

79.3%Raw benchmark value

Overall

LB · Professional reasoning · Objective

Average objective performance across LiveBench's current public category mix.

Rank #19 · Source label: gpt-5.4-xhigh

verified runtimeexact alias

Raw row drilldownsource row, percentile, last updated, eligibility

Source: LiveBench
Raw value: 78%
Percentile: 72.3%
Last updated: stale
Eligibility: headline eligible

Derived from the official LiveBench website leaderboard table. Category averages included: 7.

72.3% percentile inside its fair comparison set

78%Raw benchmark value

Consecutive events

LB · Professional reasoning · Objective

Objective consecutive events score in LiveBench.

Rank #28 · Source label: gpt-5.4-xhigh

verified runtimeexact alias

Raw row drilldownsource row, percentile, last updated, eligibility

Source: LiveBench
Raw value: 86.2%
Percentile: 58.5%
Last updated: stale
Eligibility: headline eligible

Derived from the official LiveBench website leaderboard table. Task: consecutive_events. Category: Data Analysis.

58.5% percentile inside its fair comparison set

86.2%Raw benchmark value

Table join

LB · Professional reasoning · Objective

Objective table join score in LiveBench.

Rank #12 · Source label: gpt-5.4-xhigh

verified runtimeexact alias

Raw row drilldownsource row, percentile, last updated, eligibility

Source: LiveBench
Raw value: 51.8%
Percentile: 83.1%
Last updated: stale
Eligibility: headline eligible

Derived from the official LiveBench website leaderboard table. Task: tablejoin. Category: Data Analysis.

83.1% percentile inside its fair comparison set

51.8%Raw benchmark value

Table reformat

LB · Professional reasoning · Objective

Objective table reformat score in LiveBench.

Rank #5 · Source label: gpt-5.4-xhigh

verified runtimeexact alias

Raw row drilldownsource row, percentile, last updated, eligibility

Source: LiveBench
Raw value: 100%
Percentile: 100%
Last updated: stale
Eligibility: headline eligible

Derived from the official LiveBench website leaderboard table. Task: tablereformat. Category: Data Analysis.

100% percentile inside its fair comparison set

100%Raw benchmark value

Hallucination

BB · Professional reasoning · Rubric

Accuracy and fabrication resistance under prompts that invite unsupported claims.

Rank #15 · Source label: openai/gpt-5.4

verified runtimeexact alias

Raw row drilldownsource row, percentile, last updated, eligibility

Source: BridgeBench
Raw value: 72.8%
Percentile: 58.8%
Last updated: archived
Eligibility: headline eligible

Parsed from the BridgeBench page for bridgebench-hallucination.

58.8% percentile inside its fair comparison set

72.8%Raw benchmark value

BS pushback

BB · Professional reasoning · Rubric

Resistance to confidently accepting bogus assumptions in expert-style prompts.

Rank #4 · Source label: openai/gpt-5.4

verified runtimeexact alias

Raw row drilldownsource row, percentile, last updated, eligibility

Source: BridgeBench
Raw value: 91.5%
Percentile: 88.9%
Last updated: archived
Eligibility: headline eligible

Parsed from the BridgeBench page for bridgebench-pushback.

88.9% percentile inside its fair comparison set

91.5%Raw benchmark value

Poker Agent

VALS-AI · Professional reasoning · Objective

Agent profit in poker-style strategic play.

Rank #3 · Source label: openai/gpt-5-2025-08-07

backfilledproxy backfilledBackground only

Raw row drilldownsource row, percentile, last updated, eligibility

Source: Vals AI
Raw value: 1,103.2 score
Percentile: 94.7%
Last updated: archived
Eligibility: Fallback benchmark identity is visible for context but excluded from default ranking.

Parsed from Vals AI BenchmarkView overall scores. Vals slug: poker_agent; provider: unknown. Backfilled from GPT-5 via approved benchmark identity mapping map-gpt-5-4-to-gpt-5.

94.7% percentile inside its fair comparison set

1,103.2 scoreRaw benchmark valueCI 1,103.2 score - 1,103.2 score

Search / tool use4 benchmarks54.6%

Tau2-Bench Telecom

AA · Search / tool use · Objective

It matters when the model must browse, call tools, and recover useful answers from external systems.

Rank #170 · Source label: GPT-5.4 (Non-reasoning)

verified runtimeexact alias

Raw row drilldownsource row, percentile, last updated, eligibility

Source: Artificial Analysis
Raw value: 36%
Percentile: 50.1%
Last updated: recent
Eligibility: headline eligible

Parsed from Artificial Analysis public leaderboard field `tau2`.

50.1% percentile inside its fair comparison set

36%Raw benchmark value

Search Arena

AR · Search / tool use · Human

It matters when the model must browse, call tools, and recover useful answers from external systems.

Rank #13 · Source label: gpt-5.4-search

verified runtimeexact alias

Raw row drilldownsource row, percentile, last updated, eligibility

Source: Arena
Raw value: 1,195
Percentile: 63.6%
Last updated: recent
Eligibility: headline eligible

Parsed from Arena leaderboard dataset row `gpt-5.4-search`. Category: overall. Source rank: #14. Votes: 110116. Organization: openai. License: Proprietary.

63.6% percentile inside its fair comparison set

1,195Raw benchmark valueCI 1,189 - 1,200

Search Arena · No Style Control

AR · Search / tool use · Human

It matters when the model must browse, call tools, and recover useful answers from external systems.

Rank #16 · Source label: gpt-5.4-search

verified runtimeexact alias

Raw row drilldownsource row, percentile, last updated, eligibility

Source: Arena
Raw value: 1,197
Percentile: 54.5%
Last updated: recent
Eligibility: headline eligible

Parsed from Arena leaderboard dataset row `gpt-5.4-search`. Category: overall. Source rank: #16. Votes: 110116. Organization: openai. License: Proprietary.

54.5% percentile inside its fair comparison set

1,197Raw benchmark valueCI 1,192 - 1,203

BrowseComp

OFF · Search / tool use · Objective

It matters when the model must browse, call tools, and recover useful answers from external systems.

Rank #4

Official company resultmanual verifiedmanual verified

Raw row drilldownsource row, percentile, last updated, eligibility

Source: Provider official evals
Raw value: 82.7%
Percentile: 50%
Last updated: archived
Eligibility: headline eligible

Verified against OpenAI's GPT-5.4 launch page dated March 5, 2026. Uses the published BrowseComp figure. Checked 2026-04-29. Verification: manual_public_page_verification.

50% percentile inside its fair comparison set

82.7%Raw benchmark value

Long context2 benchmarks64%

Long Context Reasoning

AA · Long context · Objective

It checks whether long-context claims survive contact with retrieval, memory, or long-document tasks.

Rank #204 · Source label: GPT-5.4 (Non-reasoning)

verified runtimeexact alias

Raw row drilldownsource row, percentile, last updated, eligibility

Source: Artificial Analysis
Raw value: 58.3%
Percentile: 55.4%
Last updated: recent
Eligibility: headline eligible

Parsed from Artificial Analysis public leaderboard field `lcr`.

55.4% percentile inside its fair comparison set

58.3%Raw benchmark value

CorpFin v2

VALS-AI · Long context · Objective

It checks whether long-context claims survive contact with retrieval, memory, or long-document tasks.

Rank #33 · Source label: openai/gpt-5.4-2026-03-05

verified runtimeexact alias

Raw row drilldownsource row, percentile, last updated, eligibility

Source: Vals AI
Raw value: 65.3%
Percentile: 72.6%
Last updated: aging
Eligibility: headline eligible

Parsed from Vals AI BenchmarkView overall scores. Vals slug: corp_fin_v2; provider: OpenAI.

72.6% percentile inside its fair comparison set

65.3%Raw benchmark valueCI 63.4% - 67.1%

Vision understanding25 benchmarks84.8%

MMMU-Pro

AA · Vision understanding · Objective

It is useful when the model must read charts, UI, screenshots, or visual scenes rather than text alone.

Rank #89 · Source label: GPT-5.4 (Non-reasoning)

verified runtimeexact alias

Raw row drilldownsource row, percentile, last updated, eligibility

Source: Artificial Analysis
Raw value: 70.6%
Percentile: 57.7%
Last updated: recent
Eligibility: headline eligible

Parsed from Artificial Analysis public leaderboard field `mmmuPro`.

57.7% percentile inside its fair comparison set

70.6%Raw benchmark value

VISTA

SL · Vision understanding · Rubric

It is useful when the model must read charts, UI, screenshots, or visual scenes rather than text alone.

Rank #3 · Source label: gpt-5.4-pro-2026-03-05

verified runtimeexact alias

Raw row drilldownsource row, percentile, last updated, eligibility

Source: Scale Labs
Raw value: 53.9%
Percentile: 96.8%
Last updated: recent
Eligibility: headline eligible

Parsed from the public Scale Labs page for scale-vista. Reported model configuration: gpt-5.4-pro-2026-03-05. Collapse policy: highest reported score per canonical model.

96.8% percentile inside its fair comparison set

53.9%Raw benchmark value

VTB

SL · Vision understanding · Rubric

It is useful when the model must read charts, UI, screenshots, or visual scenes rather than text alone.

Rank #2 · Source label: gpt-5.4-2026-03-05 (reasoning effort = high)

verified runtimeexact alias

Raw row drilldownsource row, percentile, last updated, eligibility

Source: Scale Labs
Raw value: 29.2%
Percentile: 95.8%
Last updated: recent
Eligibility: headline eligible

Parsed from the public Scale Labs page for scale-vtb. Reported model configuration: gpt-5.4-2026-03-05 (reasoning effort = high). Collapse policy: highest reported score per canonical model.

95.8% percentile inside its fair comparison set

29.2%Raw benchmark value

Vision Arena

AR · Vision understanding · Human

It is useful when the model must read charts, UI, screenshots, or visual scenes rather than text alone.

Rank #21 · Source label: gpt-5.4-high

verified runtimeexact alias

Raw row drilldownsource row, percentile, last updated, eligibility

Source: Arena
Raw value: 1,284
Percentile: 85.6%
Last updated: recent
Eligibility: headline eligible

Parsed from Arena leaderboard dataset row `gpt-5.4-high`. Category: overall. Source rank: #22. Votes: 29177. Organization: openai. License: Proprietary.

85.6% percentile inside its fair comparison set

1,284Raw benchmark valueCI 1,278 - 1,290

Vision Arena · Captioning

AR · Vision understanding · Human

It is useful when the model must read charts, UI, screenshots, or visual scenes rather than text alone.

Rank #15 · Source label: gpt-5-chat

backfilledproxy backfilledBackground only

Raw row drilldownsource row, percentile, last updated, eligibility

Source: Arena
Raw value: 1,195
Percentile: 58.1%
Last updated: recent
Eligibility: Fallback benchmark identity is visible for context but excluded from default ranking.

Parsed from Arena leaderboard dataset row `gpt-5-chat`. Category: captioning. Source rank: #15. Votes: 375. Organization: openai. License: Proprietary. Backfilled from GPT-5 via approved benchmark identity mapping map-gpt-5-4-to-gpt-5.

58.1% percentile inside its fair comparison set

1,195Raw benchmark valueCI 1,163 - 1,227

Vision Arena · Creative Writing Vision

AR · Vision understanding · Human

It is useful when the model must read charts, UI, screenshots, or visual scenes rather than text alone.

Rank #27 · Source label: gpt-5.4-high

verified runtimeexact alias

Raw row drilldownsource row, percentile, last updated, eligibility

Source: Arena
Raw value: 1,260
Percentile: 68.3%
Last updated: recent
Eligibility: headline eligible

Parsed from Arena leaderboard dataset row `gpt-5.4-high`. Category: creative_writing_vision. Source rank: #31. Votes: 1945. Organization: openai. License: Proprietary.

68.3% percentile inside its fair comparison set

1,260Raw benchmark valueCI 1,245 - 1,274

Vision Arena · Diagram

AR · Vision understanding · Human

It is useful when the model must read charts, UI, screenshots, or visual scenes rather than text alone.

Rank #13 · Source label: gpt-5.4-high

verified runtimeexact alias

Raw row drilldownsource row, percentile, last updated, eligibility

Source: Arena
Raw value: 1,321
Percentile: 87.9%
Last updated: recent
Eligibility: headline eligible

Parsed from Arena leaderboard dataset row `gpt-5.4-high`. Category: diagram. Source rank: #13. Votes: 7803. Organization: openai. License: Proprietary.

87.9% percentile inside its fair comparison set

1,321Raw benchmark valueCI 1,312 - 1,329

Vision Arena · English

AR · Vision understanding · Human

It is useful when the model must read charts, UI, screenshots, or visual scenes rather than text alone.

Rank #16 · Source label: gpt-5.4-high

verified runtimeexact alias

Raw row drilldownsource row, percentile, last updated, eligibility

Source: Arena
Raw value: 1,282
Percentile: 89.1%
Last updated: recent
Eligibility: headline eligible

Parsed from Arena leaderboard dataset row `gpt-5.4-high`. Category: english. Source rank: #16. Votes: 11572. Organization: openai. License: Proprietary.

89.1% percentile inside its fair comparison set

1,282Raw benchmark valueCI 1,274 - 1,290

Vision Arena · Entity Recognition

AR · Vision understanding · Human

It is useful when the model must read charts, UI, screenshots, or visual scenes rather than text alone.

Rank #21 · Source label: gpt-5.4

verified runtimeexact alias

Raw row drilldownsource row, percentile, last updated, eligibility

Source: Arena
Raw value: 1,215
Percentile: 50%
Last updated: recent
Eligibility: headline eligible

Parsed from Arena leaderboard dataset row `gpt-5.4`. Category: entity_recognition. Source rank: #23. Votes: 134. Organization: openai. License: Proprietary.

50% percentile inside its fair comparison set

1,215Raw benchmark valueCI 1,168 - 1,262

Vision Arena · Homework

AR · Vision understanding · Human

It is useful when the model must read charts, UI, screenshots, or visual scenes rather than text alone.

Rank #9 · Source label: gpt-5.4-high

verified runtimeexact alias

Raw row drilldownsource row, percentile, last updated, eligibility

Source: Arena
Raw value: 1,331
Percentile: 91.7%
Last updated: recent
Eligibility: headline eligible

Parsed from Arena leaderboard dataset row `gpt-5.4-high`. Category: homework. Source rank: #10. Votes: 3768. Organization: openai. License: Proprietary.

91.7% percentile inside its fair comparison set

1,331Raw benchmark valueCI 1,321 - 1,342

Vision Arena · Humor

AR · Vision understanding · Human

It is useful when the model must read charts, UI, screenshots, or visual scenes rather than text alone.

Rank #22 · Source label: gpt-5.4-high

verified runtimeexact alias

Raw row drilldownsource row, percentile, last updated, eligibility

Source: Arena
Raw value: 1,263
Percentile: 71.2%
Last updated: recent
Eligibility: headline eligible

Parsed from Arena leaderboard dataset row `gpt-5.4-high`. Category: humor. Source rank: #23. Votes: 1061. Organization: openai. License: Proprietary.

71.2% percentile inside its fair comparison set

1,263Raw benchmark valueCI 1,244 - 1,283

Vision Arena · Ocr

AR · Vision understanding · Human

It is useful when the model must read charts, UI, screenshots, or visual scenes rather than text alone.

Rank #19 · Source label: gpt-5.4-high

verified runtimeexact alias

Raw row drilldownsource row, percentile, last updated, eligibility

Source: Arena
Raw value: 1,299
Percentile: 82%
Last updated: recent
Eligibility: headline eligible

Parsed from Arena leaderboard dataset row `gpt-5.4-high`. Category: ocr. Source rank: #19. Votes: 20737. Organization: openai. License: Proprietary.

82% percentile inside its fair comparison set

1,299Raw benchmark valueCI 1,293 - 1,304

Vision Arena · No Style Control

AR · Vision understanding · Human

It is useful when the model must read charts, UI, screenshots, or visual scenes rather than text alone.

Rank #16 · Source label: gpt-5.4-high

verified runtimeexact alias

Raw row drilldownsource row, percentile, last updated, eligibility

Source: Arena
Raw value: 1,302
Percentile: 89.2%
Last updated: recent
Eligibility: headline eligible

Parsed from Arena leaderboard dataset row `gpt-5.4-high`. Category: overall. Source rank: #17. Votes: 29177. Organization: openai. License: Proprietary.

89.2% percentile inside its fair comparison set

1,302Raw benchmark valueCI 1,296 - 1,308

Vision Arena · Captioning · No Style Control

AR · Vision understanding · Human

It is useful when the model must read charts, UI, screenshots, or visual scenes rather than text alone.

Rank #19 · Source label: gpt-5-chat

backfilledproxy backfilledBackground only

Raw row drilldownsource row, percentile, last updated, eligibility

Source: Arena
Raw value: 1,212
Percentile: 45.2%
Last updated: recent
Eligibility: Fallback benchmark identity is visible for context but excluded from default ranking.

Parsed from Arena leaderboard dataset row `gpt-5-chat`. Category: captioning. Source rank: #19. Votes: 375. Organization: openai. License: Proprietary. Backfilled from GPT-5 via approved benchmark identity mapping map-gpt-5-4-to-gpt-5.

45.2% percentile inside its fair comparison set

1,212Raw benchmark valueCI 1,181 - 1,244

Vision Arena · Creative Writing Vision · No Style Control

AR · Vision understanding · Human

It is useful when the model must read charts, UI, screenshots, or visual scenes rather than text alone.

Rank #23 · Source label: gpt-5.4-high

verified runtimeexact alias

Raw row drilldownsource row, percentile, last updated, eligibility

Source: Arena
Raw value: 1,281
Percentile: 73.2%
Last updated: recent
Eligibility: headline eligible

Parsed from Arena leaderboard dataset row `gpt-5.4-high`. Category: creative_writing_vision. Source rank: #26. Votes: 1945. Organization: openai. License: Proprietary.

73.2% percentile inside its fair comparison set

1,281Raw benchmark valueCI 1,267 - 1,295

Vision Arena · Diagram · No Style Control

AR · Vision understanding · Human

It is useful when the model must read charts, UI, screenshots, or visual scenes rather than text alone.

Rank #11 · Source label: gpt-5.4-high

verified runtimeexact alias

Raw row drilldownsource row, percentile, last updated, eligibility

Source: Arena
Raw value: 1,326
Percentile: 89.9%
Last updated: recent
Eligibility: headline eligible

Parsed from Arena leaderboard dataset row `gpt-5.4-high`. Category: diagram. Source rank: #11. Votes: 7803. Organization: openai. License: Proprietary.

89.9% percentile inside its fair comparison set

1,326Raw benchmark valueCI 1,318 - 1,335

Vision Arena · English · No Style Control

AR · Vision understanding · Human

It is useful when the model must read charts, UI, screenshots, or visual scenes rather than text alone.

Rank #13 · Source label: gpt-5.4-high

verified runtimeexact alias

Raw row drilldownsource row, percentile, last updated, eligibility

Source: Arena
Raw value: 1,301
Percentile: 91.3%
Last updated: recent
Eligibility: headline eligible

Parsed from Arena leaderboard dataset row `gpt-5.4-high`. Category: english. Source rank: #14. Votes: 11572. Organization: openai. License: Proprietary.

91.3% percentile inside its fair comparison set

1,301Raw benchmark valueCI 1,293 - 1,310

Vision Arena · Entity Recognition · No Style Control

AR · Vision understanding · Human

It is useful when the model must read charts, UI, screenshots, or visual scenes rather than text alone.

Rank #15 · Source label: gpt-5.4

verified runtimeexact alias

Raw row drilldownsource row, percentile, last updated, eligibility

Source: Arena
Raw value: 1,251
Percentile: 65%
Last updated: recent
Eligibility: headline eligible

Parsed from Arena leaderboard dataset row `gpt-5.4`. Category: entity_recognition. Source rank: #17. Votes: 134. Organization: openai. License: Proprietary.

65% percentile inside its fair comparison set

1,251Raw benchmark valueCI 1,205 - 1,298

Vision Arena · Homework · No Style Control

AR · Vision understanding · Human

It is useful when the model must read charts, UI, screenshots, or visual scenes rather than text alone.

Rank #11 · Source label: gpt-5.4-high

verified runtimeexact alias

Raw row drilldownsource row, percentile, last updated, eligibility

Source: Arena
Raw value: 1,337
Percentile: 89.6%
Last updated: recent
Eligibility: headline eligible

Parsed from Arena leaderboard dataset row `gpt-5.4-high`. Category: homework. Source rank: #11. Votes: 3768. Organization: openai. License: Proprietary.

89.6% percentile inside its fair comparison set

1,337Raw benchmark valueCI 1,327 - 1,348

Vision Arena · Humor · No Style Control

AR · Vision understanding · Human

It is useful when the model must read charts, UI, screenshots, or visual scenes rather than text alone.

Rank #15 · Source label: gpt-5.4-high

verified runtimeexact alias

Raw row drilldownsource row, percentile, last updated, eligibility

Source: Arena
Raw value: 1,284
Percentile: 80.8%
Last updated: recent
Eligibility: headline eligible

Parsed from Arena leaderboard dataset row `gpt-5.4-high`. Category: humor. Source rank: #16. Votes: 1061. Organization: openai. License: Proprietary.

80.8% percentile inside its fair comparison set

1,284Raw benchmark valueCI 1,265 - 1,303

Vision Arena · Ocr · No Style Control

AR · Vision understanding · Human

It is useful when the model must read charts, UI, screenshots, or visual scenes rather than text alone.

Rank #14 · Source label: gpt-5.4-high

verified runtimeexact alias

Raw row drilldownsource row, percentile, last updated, eligibility

Source: Arena
Raw value: 1,311
Percentile: 87%
Last updated: recent
Eligibility: headline eligible

Parsed from Arena leaderboard dataset row `gpt-5.4-high`. Category: ocr. Source rank: #15. Votes: 20737. Organization: openai. License: Proprietary.

87% percentile inside its fair comparison set

1,311Raw benchmark valueCI 1,305 - 1,317

MMMU Pro

VALS-AI · Vision understanding · Objective

It is useful when the model must read charts, UI, screenshots, or visual scenes rather than text alone.

Rank #15 · Source label: openai/gpt-5.4-2026-03-05

verified runtimeexact alias

Raw row drilldownsource row, percentile, last updated, eligibility

Source: Vals AI
Raw value: 87.5%
Percentile: 84.3%
Last updated: recent
Eligibility: headline eligible

Parsed from Vals AI BenchmarkView overall scores. Vals slug: mmmu; provider: OpenAI.

84.3% percentile inside its fair comparison set

87.5%Raw benchmark valueCI 86% - 89.1%

MMMU-Pro

OFF · Vision understanding · Objective

It is useful when the model must read charts, UI, screenshots, or visual scenes rather than text alone.

Rank #2

Official company resultmanual verifiedmanual verified

Raw row drilldownsource row, percentile, last updated, eligibility

Source: Provider official evals
Raw value: 81.2%
Percentile: 100%
Last updated: archived
Eligibility: headline eligible

Verified against OpenAI's GPT-5.4 launch page dated March 5, 2026. Uses the published no-tools MMMU Pro figure. Checked 2026-04-29. Verification: manual_public_page_verification.

100% percentile inside its fair comparison set

81.2%Raw benchmark value

Vision Arena · Creative Writing

AR · Vision understanding · Human

It is useful when the model must read charts, UI, screenshots, or visual scenes rather than text alone.

Rank #12 · Source label: gpt-5-chat

backfilledproxy backfilledBackground only

Raw row drilldownsource row, percentile, last updated, eligibility

Source: Arena
Raw value: 1,232
Percentile: 68.8%
Last updated: archived
Eligibility: Fallback benchmark identity is visible for context but excluded from default ranking.

Parsed from Arena leaderboard dataset row `gpt-5-chat`. Category: creative_writing. Source rank: #12. Votes: 1518. Organization: openai. License: Proprietary. Backfilled from GPT-5 via approved benchmark identity mapping map-gpt-5-4-to-gpt-5.

68.8% percentile inside its fair comparison set

1,232Raw benchmark valueCI 1,216 - 1,248

Vision Arena · Creative Writing · No Style Control

AR · Vision understanding · Human

It is useful when the model must read charts, UI, screenshots, or visual scenes rather than text alone.

Rank #11 · Source label: gpt-5-chat

backfilledproxy backfilledBackground only

Raw row drilldownsource row, percentile, last updated, eligibility

Source: Arena
Raw value: 1,239
Percentile: 71.9%
Last updated: archived
Eligibility: Fallback benchmark identity is visible for context but excluded from default ranking.

71.9% percentile inside its fair comparison set

1,239Raw benchmark valueCI 1,224 - 1,255

Document understanding4 benchmarks70.7%

Document Arena

AR · Document understanding · Human

It matters when the job is reading PDFs, tables, forms, or mixed-layout documents rather than plain chat.

Rank #17 · Source label: gpt-5.4

verified runtimeexact alias

Raw row drilldownsource row, percentile, last updated, eligibility

Source: Arena
Raw value: 1,470
Percentile: 60%
Last updated: recent
Eligibility: headline eligible

Parsed from Arena leaderboard dataset row `gpt-5.4`. Category: overall. Source rank: #18. Votes: 33331. Organization: openai. License: Proprietary.

60% percentile inside its fair comparison set

1,470Raw benchmark valueCI 1,464 - 1,476

Document Arena · No Style Control

AR · Document understanding · Human

It matters when the job is reading PDFs, tables, forms, or mixed-layout documents rather than plain chat.

Rank #13 · Source label: gpt-5.4

verified runtimeexact alias

Raw row drilldownsource row, percentile, last updated, eligibility

Source: Arena
Raw value: 1,471
Percentile: 70%
Last updated: recent
Eligibility: headline eligible

Parsed from Arena leaderboard dataset row `gpt-5.4`. Category: overall. Source rank: #14. Votes: 33331. Organization: openai. License: Proprietary.

70% percentile inside its fair comparison set

1,471Raw benchmark valueCI 1,465 - 1,477

MortgageTax

VALS-AI · Document understanding · Objective

It matters when the job is reading PDFs, tables, forms, or mixed-layout documents rather than plain chat.

Rank #17 · Source label: openai/gpt-5.4-2026-03-05

verified runtimeexact alias

Raw row drilldownsource row, percentile, last updated, eligibility

Source: Vals AI
Raw value: 68.3%
Percentile: 81.4%
Last updated: recent
Eligibility: headline eligible

Parsed from Vals AI BenchmarkView overall scores. Vals slug: mortgage_tax; provider: OpenAI.

81.4% percentile inside its fair comparison set

68.3%Raw benchmark valueCI 66.5% - 70.1%

Multimodal mix

OC · Document understanding · Objective

It matters when the job is reading PDFs, tables, forms, or mixed-layout documents rather than plain chat.

Rank #3 · Source label: gpt-5

backfilledproxy backfilledBackground only

Raw row drilldownsource row, percentile, last updated, eligibility

Source: OpenCompass
Raw value: 75.4%
Percentile: 92.9%
Last updated: archived
Eligibility: Fallback benchmark identity is visible for context but excluded from default ranking.

Backfilled from GPT-5 via approved benchmark identity mapping map-gpt-5-4-to-gpt-5.

92.9% percentile inside its fair comparison set

75.4%Raw benchmark value

Safety1 benchmark91%

MASK

SL · Safety · Rubric

Whether a model stays honest instead of covertly optimizing against the user.

Rank #7 · Source label: gpt-5.4-pro-2026-03-05

verified runtimeexact alias

Raw row drilldownsource row, percentile, last updated, eligibility

Source: Scale Labs
Raw value: 91.7%
Percentile: 91%
Last updated: recent
Eligibility: headline eligible

Parsed from the public Scale Labs page for scale-mask. Reported model configuration: gpt-5.4-pro-2026-03-05. Collapse policy: highest reported score per canonical model.

91% percentile inside its fair comparison set

91.7%Raw benchmark value

Embeddings / retrieval1 benchmarkn/a

Retrieval

MTEB · Embeddings / retrieval · Retrieval

It is one of the few direct signals for retrieval stacks, where embedding quality matters more than chat style.

Rank #2 · Source label: gpt-5

backfilledproxy backfilledBackground only

Raw row drilldownsource row, percentile, last updated, eligibility

Source: MTEB
Raw value: 58.8 ndcg
Percentile: 100%
Last updated: archived
Eligibility: Fallback benchmark identity is visible for context but excluded from default ranking.

Embedding endpoint score. Backfilled from GPT-5 via approved benchmark identity mapping map-gpt-5-4-to-gpt-5.

100% percentile inside its fair comparison set

58.8 ndcgRaw benchmark value

Multilingual20 benchmarksn/a

Text Arena · Chinese

AR · Multilingual · Human

Observed user preference in Arena's Text Arena chinese leaderboard.

Rank #46 · Source label: gpt-5.4-high

verified runtimeexact alias

Raw row drilldownsource row, percentile, last updated, eligibility

Source: Arena
Raw value: 1,507
Percentile: 87.2%
Last updated: recent
Eligibility: headline eligible

Parsed from Arena leaderboard dataset row `gpt-5.4-high`. Category: chinese. Source rank: #48. Votes: 4132. Organization: openai. License: Proprietary.

87.2% percentile inside its fair comparison set

1,507Raw benchmark valueCI 1,497 - 1,517

Text Arena · French

AR · Multilingual · Human

Observed user preference in Arena's Text Arena french leaderboard.

Rank #24 · Source label: gpt-5.4-high

verified runtimeexact alias

Raw row drilldownsource row, percentile, last updated, eligibility

Source: Arena
Raw value: 1,500
Percentile: 91.1%
Last updated: recent
Eligibility: headline eligible

Parsed from Arena leaderboard dataset row `gpt-5.4-high`. Category: french. Source rank: #24. Votes: 2396. Organization: openai. License: Proprietary.

91.1% percentile inside its fair comparison set

1,500Raw benchmark valueCI 1,486 - 1,514

Text Arena · German

AR · Multilingual · Human

Observed user preference in Arena's Text Arena german leaderboard.

Rank #31 · Source label: gpt-5.4-high

verified runtimeexact alias

Raw row drilldownsource row, percentile, last updated, eligibility

Source: Arena
Raw value: 1,476
Percentile: 89%
Last updated: recent
Eligibility: headline eligible

Parsed from Arena leaderboard dataset row `gpt-5.4-high`. Category: german. Source rank: #33. Votes: 1088. Organization: openai. License: Proprietary.

89% percentile inside its fair comparison set

1,476Raw benchmark valueCI 1,457 - 1,495

Text Arena · Japanese

AR · Multilingual · Human

Observed user preference in Arena's Text Arena japanese leaderboard.

Rank #19 · Source label: gpt-5.4-high

verified runtimeexact alias

Raw row drilldownsource row, percentile, last updated, eligibility

Source: Arena
Raw value: 1,479
Percentile: 92.6%
Last updated: recent
Eligibility: headline eligible

Parsed from Arena leaderboard dataset row `gpt-5.4-high`. Category: japanese. Source rank: #19. Votes: 671. Organization: openai. License: Proprietary.

92.6% percentile inside its fair comparison set

1,479Raw benchmark valueCI 1,454 - 1,504

Text Arena · Korean

AR · Multilingual · Human

Observed user preference in Arena's Text Arena korean leaderboard.

Rank #34 · Source label: gpt-5.4-high

verified runtimeexact alias

Raw row drilldownsource row, percentile, last updated, eligibility

Source: Arena
Raw value: 1,435
Percentile: 86.5%
Last updated: recent
Eligibility: headline eligible

Parsed from Arena leaderboard dataset row `gpt-5.4-high`. Category: korean. Source rank: #36. Votes: 1106. Organization: openai. License: Proprietary.

86.5% percentile inside its fair comparison set

1,435Raw benchmark valueCI 1,415 - 1,454

Text Arena · Non English

AR · Multilingual · Human

Observed user preference in Arena's Text Arena non_english leaderboard.

Rank #29 · Source label: gpt-5.4-high

verified runtimeexact alias

Raw row drilldownsource row, percentile, last updated, eligibility

Source: Arena
Raw value: 1,468
Percentile: 92.6%
Last updated: recent
Eligibility: headline eligible

Parsed from Arena leaderboard dataset row `gpt-5.4-high`. Category: non_english. Source rank: #29. Votes: 35418. Organization: openai. License: Proprietary.

92.6% percentile inside its fair comparison set

1,468Raw benchmark valueCI 1,464 - 1,473

Text Arena · Polish

AR · Multilingual · Human

Observed user preference in Arena's Text Arena polish leaderboard.

Rank #25 · Source label: gpt-5.4-high

verified runtimeexact alias

Raw row drilldownsource row, percentile, last updated, eligibility

Source: Arena
Raw value: 1,482
Percentile: 87.9%
Last updated: recent
Eligibility: headline eligible

Parsed from Arena leaderboard dataset row `gpt-5.4-high`. Category: polish. Source rank: #26. Votes: 1320. Organization: openai. License: Proprietary.

87.9% percentile inside its fair comparison set

1,482Raw benchmark valueCI 1,465 - 1,498

Text Arena · Russian

AR · Multilingual · Human

Observed user preference in Arena's Text Arena russian leaderboard.

Rank #23 · Source label: gpt-5.4-high

verified runtimeexact alias

Raw row drilldownsource row, percentile, last updated, eligibility

Source: Arena
Raw value: 1,487
Percentile: 93.5%
Last updated: recent
Eligibility: headline eligible

Parsed from Arena leaderboard dataset row `gpt-5.4-high`. Category: russian. Source rank: #24. Votes: 7209. Organization: openai. License: Proprietary.

93.5% percentile inside its fair comparison set

1,487Raw benchmark valueCI 1,479 - 1,495

Text Arena · Spanish

AR · Multilingual · Human

Observed user preference in Arena's Text Arena spanish leaderboard.

Rank #45 · Source label: gpt-5.4

verified runtimeexact alias

Raw row drilldownsource row, percentile, last updated, eligibility

Source: Arena
Raw value: 1,461
Percentile: 83%
Last updated: recent
Eligibility: headline eligible

Parsed from Arena leaderboard dataset row `gpt-5.4`. Category: spanish. Source rank: #49. Votes: 2037. Organization: openai. License: Proprietary.

83% percentile inside its fair comparison set

1,461Raw benchmark valueCI 1,446 - 1,475

Text Arena · Chinese · No Style Control

AR · Multilingual · Human

Observed user preference in Arena's Text Arena chinese leaderboard.

Rank #30 · Source label: gpt-5.4-high

verified runtimeexact alias

Raw row drilldownsource row, percentile, last updated, eligibility

Source: Arena
Raw value: 1,520
Percentile: 91.8%
Last updated: recent
Eligibility: headline eligible

Parsed from Arena leaderboard dataset row `gpt-5.4-high`. Category: chinese. Source rank: #32. Votes: 4132. Organization: openai. License: Proprietary.

91.8% percentile inside its fair comparison set

1,520Raw benchmark valueCI 1,510 - 1,530

Text Arena · French · No Style Control

AR · Multilingual · Human

Observed user preference in Arena's Text Arena french leaderboard.

Rank #18 · Source label: gpt-5.4-high

verified runtimeexact alias

Raw row drilldownsource row, percentile, last updated, eligibility

Source: Arena
Raw value: 1,493
Percentile: 93.4%
Last updated: recent
Eligibility: headline eligible

Parsed from Arena leaderboard dataset row `gpt-5.4-high`. Category: french. Source rank: #18. Votes: 2396. Organization: openai. License: Proprietary.

93.4% percentile inside its fair comparison set

1,493Raw benchmark valueCI 1,478 - 1,507

Text Arena · German · No Style Control

AR · Multilingual · Human

Observed user preference in Arena's Text Arena german leaderboard.

Rank #24 · Source label: gpt-5.4-high

verified runtimeexact alias

Raw row drilldownsource row, percentile, last updated, eligibility

Source: Arena
Raw value: 1,472
Percentile: 91.5%
Last updated: recent
Eligibility: headline eligible

Parsed from Arena leaderboard dataset row `gpt-5.4-high`. Category: german. Source rank: #25. Votes: 1088. Organization: openai. License: Proprietary.

91.5% percentile inside its fair comparison set

1,472Raw benchmark valueCI 1,453 - 1,491

Text Arena · Japanese · No Style Control

AR · Multilingual · Human

Observed user preference in Arena's Text Arena japanese leaderboard.

Rank #16 · Source label: gpt-5.4-high

verified runtimeexact alias

Raw row drilldownsource row, percentile, last updated, eligibility

Source: Arena
Raw value: 1,483
Percentile: 93.8%
Last updated: recent
Eligibility: headline eligible

Parsed from Arena leaderboard dataset row `gpt-5.4-high`. Category: japanese. Source rank: #16. Votes: 671. Organization: openai. License: Proprietary.

93.8% percentile inside its fair comparison set

1,483Raw benchmark valueCI 1,459 - 1,508

Text Arena · Korean · No Style Control

AR · Multilingual · Human

Observed user preference in Arena's Text Arena korean leaderboard.

Rank #19 · Source label: gpt-5.4-high

verified runtimeexact alias

Raw row drilldownsource row, percentile, last updated, eligibility

Source: Arena
Raw value: 1,450
Percentile: 92.6%
Last updated: recent
Eligibility: headline eligible

Parsed from Arena leaderboard dataset row `gpt-5.4-high`. Category: korean. Source rank: #19. Votes: 1106. Organization: openai. License: Proprietary.

92.6% percentile inside its fair comparison set

1,450Raw benchmark valueCI 1,430 - 1,469

Text Arena · Non English · No Style Control

AR · Multilingual · Human

Observed user preference in Arena's Text Arena non_english leaderboard.

Rank #25 · Source label: gpt-5.4-high

verified runtimeexact alias

Raw row drilldownsource row, percentile, last updated, eligibility

Source: Arena
Raw value: 1,465
Percentile: 93.6%
Last updated: recent
Eligibility: headline eligible

Parsed from Arena leaderboard dataset row `gpt-5.4-high`. Category: non_english. Source rank: #26. Votes: 35418. Organization: openai. License: Proprietary.

93.6% percentile inside its fair comparison set

1,465Raw benchmark valueCI 1,460 - 1,470

Text Arena · Polish · No Style Control

AR · Multilingual · Human

Observed user preference in Arena's Text Arena polish leaderboard.

Rank #25 · Source label: gpt-5.4-high

verified runtimeexact alias

Raw row drilldownsource row, percentile, last updated, eligibility

Source: Arena
Raw value: 1,471
Percentile: 87.9%
Last updated: recent
Eligibility: headline eligible

Parsed from Arena leaderboard dataset row `gpt-5.4-high`. Category: polish. Source rank: #26. Votes: 1320. Organization: openai. License: Proprietary.

87.9% percentile inside its fair comparison set

1,471Raw benchmark valueCI 1,455 - 1,488

Text Arena · Russian · No Style Control

AR · Multilingual · Human

Observed user preference in Arena's Text Arena russian leaderboard.

Rank #22 · Source label: gpt-5.4-high

verified runtimeexact alias

Raw row drilldownsource row, percentile, last updated, eligibility

Source: Arena
Raw value: 1,480
Percentile: 93.8%
Last updated: recent
Eligibility: headline eligible

Parsed from Arena leaderboard dataset row `gpt-5.4-high`. Category: russian. Source rank: #23. Votes: 7209. Organization: openai. License: Proprietary.

93.8% percentile inside its fair comparison set

1,480Raw benchmark valueCI 1,473 - 1,488

Text Arena · Spanish · No Style Control

AR · Multilingual · Human

Observed user preference in Arena's Text Arena spanish leaderboard.

Rank #43 · Source label: gpt-5.4-high

verified runtimeexact alias

Raw row drilldownsource row, percentile, last updated, eligibility

Source: Arena
Raw value: 1,454
Percentile: 83.8%
Last updated: recent
Eligibility: headline eligible

Parsed from Arena leaderboard dataset row `gpt-5.4-high`. Category: spanish. Source rank: #45. Votes: 2062. Organization: openai. License: Proprietary.

83.8% percentile inside its fair comparison set

1,454Raw benchmark valueCI 1,440 - 1,468

Vision Arena · Chinese

AR · Multilingual · Human

Observed user preference in Arena's Vision Arena chinese leaderboard.

Rank #25 · Source label: gpt-5.4-high

verified runtimeexact alias

Raw row drilldownsource row, percentile, last updated, eligibility

Source: Arena
Raw value: 1,316
Percentile: 76.9%
Last updated: recent
Eligibility: headline eligible

Parsed from Arena leaderboard dataset row `gpt-5.4-high`. Category: chinese. Source rank: #27. Votes: 1768. Organization: openai. License: Proprietary.

76.9% percentile inside its fair comparison set

1,316Raw benchmark valueCI 1,299 - 1,333

Vision Arena · Chinese · No Style Control

AR · Multilingual · Human

Observed user preference in Arena's Vision Arena chinese leaderboard.

Rank #17 · Source label: gpt-5.4-high

verified runtimeexact alias

Raw row drilldownsource row, percentile, last updated, eligibility

Source: Arena
Raw value: 1,353
Percentile: 84.6%
Last updated: recent
Eligibility: headline eligible

Parsed from Arena leaderboard dataset row `gpt-5.4-high`. Category: chinese. Source rank: #19. Votes: 1768. Organization: openai. License: Proprietary.

84.6% percentile inside its fair comparison set

1,353Raw benchmark valueCI 1,336 - 1,370

Source links and registry checks

official

Arena

Oct 3, 2026

source →

official

Arena

Oct 3, 2026

source →

official

Arena

Oct 3, 2026

source →

official

Arena

Oct 3, 2026

source →

official

Arena

Oct 3, 2026

source →

official

Arena

Oct 3, 2026

source →

official

Arena

Oct 3, 2026

source →

official

Arena

Oct 3, 2026

source →

official

Arena

Oct 3, 2026

source →

official

Arena

Oct 3, 2026

source →

official

Arena

Oct 3, 2026

source →

official

Arena

Oct 3, 2026

source →

official

Arena

Oct 3, 2026

source →

official

Arena

Oct 3, 2026

source →

official

Arena

Oct 3, 2026

source →

official

Arena

Oct 3, 2026

source →

official

Arena

Oct 3, 2026

source →

official

Arena

Oct 3, 2026

source →

official

Arena

Oct 3, 2026

source →

official

Arena

Oct 3, 2026

source →

official

Arena

Oct 3, 2026

source →

official

Arena

Oct 3, 2026

source →

official

Arena

Oct 3, 2026

source →

official

Arena

Oct 3, 2026

source →

official

Arena

Oct 3, 2026

source →

official

Arena

Oct 3, 2026

source →

official

Arena

Oct 3, 2026

source →

official

Arena

Oct 3, 2026

source →

official

Arena

Oct 3, 2026

source →

official

Arena

Oct 3, 2026

source →

official

Arena

Oct 3, 2026

source →

official

Arena

Oct 3, 2026

source →

official

Arena

Oct 3, 2026

source →

official

Arena

Oct 3, 2026

source →

official

Arena

Oct 3, 2026

source →

official

Arena

Oct 3, 2026

source →

official

Arena

Oct 3, 2026

source →

official

Arena

Oct 3, 2026

source →

official

Arena

Oct 3, 2026

source →

official

Arena

Oct 3, 2026

source →

official

Arena

Oct 3, 2026

source →

official

Arena

Oct 3, 2026

source →

official

Arena

Oct 3, 2026

source →

official

Arena

Oct 3, 2026

source →

official

Artificial Analysis

Oct 3, 2026

source →

official

LiveBench

Oct 3, 2026

source →

official

Scale Labs

Oct 3, 2026

source →

official

Scale Labs

Oct 3, 2026

source →

official

Scale Labs

Oct 3, 2026

source →

official

Scale Labs

Oct 3, 2026

source →

official

Scale Labs

Oct 3, 2026

source →

official

Scale Labs

Oct 3, 2026

source →

official

Scale Labs

Oct 3, 2026

source →

Model profile · OpenAI

GPT-5.4

Closed weightsfrontier · registry tag 2026 flagship

Verified but aging

Reads as verified but aging across the resolved source data. Capability fit excludes speed and cost.

Visible coverage: 66.4%
Verified coverage: 64.4%
Spread: 87.5%
Last verified: Oct 3, 2026

69%capability fit

textcodevisiondocumentsearch19 aliases53 official source links

Open compare

Data version

Current snapshot.

Published Oct 3, 2026Model list checked4 stale sources105 claim warnings9 providers · 1334 tracked models

The registry snapshot and page stamp are shown so a stale deploy is visible at a glance. View live data status.

Source-linked scores by benchmark

Verified but agingThis model currently reads as verified but aging across the resolved source data.

Chat / text36 benchmarks69.8%

Intelligence Index

AA · Chat / text · Combined

It tests whether the model is actually useful in normal conversational turns, not just on narrow correctness tasks.

Rank #182 · Source label: GPT-5.4 (Non-reasoning)

verified runtimeexact alias

Raw row drilldownsource row, percentile, last updated, eligibility

Source: Artificial Analysis
Raw value: 18
Percentile: 67.4%
Last updated: recent
Eligibility: headline eligible

Parsed from Artificial Analysis public leaderboard field `intelligenceIndex`.

67.4% percentile inside its fair comparison set

18Raw benchmark value

AA-Omniscience accuracy

AA · Chat / text · Objective

It tests whether the model is actually useful in normal conversational turns, not just on narrow correctness tasks.

Rank #97 · Source label: GPT-5.4 (Non-reasoning)

verified runtimeexact alias

Raw row drilldownsource row, percentile, last updated, eligibility

Source: Artificial Analysis
Raw value: 37.4%
Percentile: 78.8%
Last updated: recent
Eligibility: headline eligible

Parsed from Artificial Analysis public leaderboard field `omniscienceAccuracy`.

78.8% percentile inside its fair comparison set

37.4%Raw benchmark value

AA-Omniscience non-hallucination

AA · Chat / text · Objective

It tests whether the model is actually useful in normal conversational turns, not just on narrow correctness tasks.

Rank #364 · Source label: GPT-5.4 (Xhigh)

verified runtimeexact alias

Raw row drilldownsource row, percentile, last updated, eligibility

Source: Artificial Analysis
Raw value: 8.3%
Percentile: 19.9%
Last updated: recent
Eligibility: headline eligible

Parsed from Artificial Analysis public leaderboard field `omniscienceNonHallucination`.

19.9% percentile inside its fair comparison set

8.3%Raw benchmark value

IFBench

AA · Chat / text · Objective

It tests whether the model is actually useful in normal conversational turns, not just on narrow correctness tasks.

Rank #116 · Source label: GPT-5.4 (Non-reasoning)

verified runtimeexact alias

Raw row drilldownsource row, percentile, last updated, eligibility

Source: Artificial Analysis
Raw value: 48.4%
Percentile: 67%
Last updated: recent
Eligibility: headline eligible

Parsed from Artificial Analysis public leaderboard field `ifbench`.

67% percentile inside its fair comparison set

48.4%Raw benchmark value

Input price

AA · Chat / text · Speed / cost

It tests whether the model is actually useful in normal conversational turns, not just on narrow correctness tasks.

Rank #304 · Source label: GPT-5.4 (Xhigh)

verified runtimeexact alias

Raw row drilldownsource row, percentile, last updated, eligibility

Source: Artificial Analysis
Raw value: $2.5 /1M input tokens
Percentile: 14.7%
Last updated: recent
Eligibility: headline eligible

Parsed from Artificial Analysis public leaderboard field `price1mInputTokens`.

14.7% percentile inside its fair comparison set

$2.5 /1M input tokensRaw benchmark value

Output price

AA · Chat / text · Speed / cost

It tests whether the model is actually useful in normal conversational turns, not just on narrow correctness tasks.

Rank #309 · Source label: GPT-5.4 (Xhigh)

verified runtimeexact alias

Raw row drilldownsource row, percentile, last updated, eligibility

Source: Artificial Analysis
Raw value: $15 /1M output tokens
Percentile: 13.8%
Last updated: recent
Eligibility: headline eligible

Parsed from Artificial Analysis public leaderboard field `price1mOutputTokens`.

13.8% percentile inside its fair comparison set

$15 /1M output tokensRaw benchmark value

Output Speed

AA · Chat / text · Speed / cost

It tests whether the model is actually useful in normal conversational turns, not just on narrow correctness tasks.

Rank #190 · Source label: GPT-5.4 (Low)

verified runtimeexact alias

Raw row drilldownsource row, percentile, last updated, eligibility

Source: Artificial Analysis
Raw value: 59.7 tokens/s
Percentile: 27.9%
Last updated: recent
Eligibility: headline eligible

Parsed from Artificial Analysis public leaderboard field `medianOutputTokensPerSecond`.

27.9% percentile inside its fair comparison set

59.7 tokens/sRaw benchmark value

Time to first token

AA · Chat / text · Speed / cost

It tests whether the model is actually useful in normal conversational turns, not just on narrow correctness tasks.

Rank #260 · Source label: GPT-5.4 (Xhigh)

verified runtimeexact alias

Raw row drilldownsource row, percentile, last updated, eligibility

Source: Artificial Analysis
Raw value: 192.94s
Percentile: 1.9%
Last updated: recent
Eligibility: headline eligible

Parsed from Artificial Analysis public leaderboard field `medianTimeToFirstTokenSeconds`.

1.9% percentile inside its fair comparison set

192.94sRaw benchmark value

Time to first answer token

AA · Chat / text · Speed / cost

It tests whether the model is actually useful in normal conversational turns, not just on narrow correctness tasks.

Rank #258 · Source label: GPT-5.4 (Xhigh)

verified runtimeexact alias

Raw row drilldownsource row, percentile, last updated, eligibility

Source: Artificial Analysis
Raw value: 192.94s
Percentile: 1.9%
Last updated: recent
Eligibility: headline eligible

Parsed from Artificial Analysis public leaderboard field `medianTimeToFirstAnswerTokenSeconds`.

1.9% percentile inside its fair comparison set

192.94sRaw benchmark value

Text Arena

AR · Chat / text · Human

It tests whether the model is actually useful in normal conversational turns, not just on narrow correctness tasks.

Rank #32 · Source label: gpt-5.4-high

verified runtimeexact alias

Raw row drilldownsource row, percentile, last updated, eligibility

Source: Arena
Raw value: 1,475
Percentile: 91.8%
Last updated: recent
Eligibility: headline eligible

Parsed from Arena leaderboard dataset row `gpt-5.4-high`. Category: overall. Source rank: #34. Votes: 64642. Organization: openai. License: Proprietary.

91.8% percentile inside its fair comparison set

1,475Raw benchmark valueCI 1,471 - 1,479

Text Arena · Creative Writing

AR · Chat / text · Human

It tests whether the model is actually useful in normal conversational turns, not just on narrow correctness tasks.

Rank #50 · Source label: gpt-5.4-high

verified runtimeexact alias

Raw row drilldownsource row, percentile, last updated, eligibility

Source: Arena
Raw value: 1,442
Percentile: 86.9%
Last updated: recent
Eligibility: headline eligible

Parsed from Arena leaderboard dataset row `gpt-5.4-high`. Category: creative_writing. Source rank: #55. Votes: 11109. Organization: openai. License: Proprietary.

86.9% percentile inside its fair comparison set

1,442Raw benchmark valueCI 1,436 - 1,449

Text Arena · English

AR · Chat / text · Human

It tests whether the model is actually useful in normal conversational turns, not just on narrow correctness tasks.

Rank #44 · Source label: gpt-5.4-high

verified runtimeexact alias

Raw row drilldownsource row, percentile, last updated, eligibility

Source: Arena
Raw value: 1,476
Percentile: 88.6%
Last updated: recent
Eligibility: headline eligible

Parsed from Arena leaderboard dataset row `gpt-5.4-high`. Category: english. Source rank: #47. Votes: 29222. Organization: openai. License: Proprietary.

88.6% percentile inside its fair comparison set

1,476Raw benchmark valueCI 1,471 - 1,481

Text Arena · Exclude Ties

AR · Chat / text · Human

It tests whether the model is actually useful in normal conversational turns, not just on narrow correctness tasks.

Rank #36 · Source label: gpt-5.4-high

verified runtimeexact alias

Raw row drilldownsource row, percentile, last updated, eligibility

Source: Arena
Raw value: 1,484
Percentile: 90.7%
Last updated: recent
Eligibility: headline eligible

Parsed from Arena leaderboard dataset row `gpt-5.4-high`. Category: exclude_ties. Source rank: #38. Votes: 49582. Organization: openai. License: Proprietary.

90.7% percentile inside its fair comparison set

1,484Raw benchmark valueCI 1,479 - 1,489

Text Arena · Hard Prompts

AR · Chat / text · Human

It tests whether the model is actually useful in normal conversational turns, not just on narrow correctness tasks.

Rank #36 · Source label: gpt-5.4-high

verified runtimeexact alias

Raw row drilldownsource row, percentile, last updated, eligibility

Source: Arena
Raw value: 1,497
Percentile: 90.7%
Last updated: recent
Eligibility: headline eligible

Parsed from Arena leaderboard dataset row `gpt-5.4-high`. Category: hard_prompts. Source rank: #37. Votes: 42068. Organization: openai. License: Proprietary.

90.7% percentile inside its fair comparison set

1,497Raw benchmark valueCI 1,492 - 1,501

Text Arena · Hard Prompts English

AR · Chat / text · Human

It tests whether the model is actually useful in normal conversational turns, not just on narrow correctness tasks.

Rank #45 · Source label: gpt-5.4-high

verified runtimeexact alias

Raw row drilldownsource row, percentile, last updated, eligibility

Source: Arena
Raw value: 1,493
Percentile: 88.2%
Last updated: recent
Eligibility: headline eligible

Parsed from Arena leaderboard dataset row `gpt-5.4-high`. Category: hard_prompts_english. Source rank: #46. Votes: 19665. Organization: openai. License: Proprietary.

88.2% percentile inside its fair comparison set

1,493Raw benchmark valueCI 1,487 - 1,499

Text Arena · Instruction Following

AR · Chat / text · Human

It tests whether the model is actually useful in normal conversational turns, not just on narrow correctness tasks.

Rank #35 · Source label: gpt-5.4-high

verified runtimeexact alias

Raw row drilldownsource row, percentile, last updated, eligibility

Source: Arena
Raw value: 1,471
Percentile: 91%
Last updated: recent
Eligibility: headline eligible

Parsed from Arena leaderboard dataset row `gpt-5.4-high`. Category: instruction_following. Source rank: #36. Votes: 21838. Organization: openai. License: Proprietary.

91% percentile inside its fair comparison set

1,471Raw benchmark valueCI 1,466 - 1,477

Text Arena · Longer Query

AR · Chat / text · Human

It tests whether the model is actually useful in normal conversational turns, not just on narrow correctness tasks.

Rank #42 · Source label: gpt-5.4-high

verified runtimeexact alias

Raw row drilldownsource row, percentile, last updated, eligibility

Source: Arena
Raw value: 1,481
Percentile: 88.5%
Last updated: recent
Eligibility: headline eligible

Parsed from Arena leaderboard dataset row `gpt-5.4-high`. Category: longer_query. Source rank: #43. Votes: 28402. Organization: openai. License: Proprietary.

88.5% percentile inside its fair comparison set

1,481Raw benchmark valueCI 1,476 - 1,487

Text Arena · Multi Turn

AR · Chat / text · Human

It tests whether the model is actually useful in normal conversational turns, not just on narrow correctness tasks.

Rank #18 · Source label: gpt-5.4-high

verified runtimeexact alias

Raw row drilldownsource row, percentile, last updated, eligibility

Source: Arena
Raw value: 1,493
Percentile: 95.5%
Last updated: recent
Eligibility: headline eligible

Parsed from Arena leaderboard dataset row `gpt-5.4-high`. Category: multi_turn. Source rank: #18. Votes: 11988. Organization: openai. License: Proprietary.

95.5% percentile inside its fair comparison set

1,493Raw benchmark valueCI 1,486 - 1,500

Text Arena · No Style Control

AR · Chat / text · Human

It tests whether the model is actually useful in normal conversational turns, not just on narrow correctness tasks.

Rank #30 · Source label: gpt-5.4-high

verified runtimeexact alias

Raw row drilldownsource row, percentile, last updated, eligibility

Source: Arena
Raw value: 1,469
Percentile: 92.3%
Last updated: recent
Eligibility: headline eligible

Parsed from Arena leaderboard dataset row `gpt-5.4-high`. Category: overall. Source rank: #31. Votes: 64642. Organization: openai. License: Proprietary.

92.3% percentile inside its fair comparison set

1,469Raw benchmark valueCI 1,465 - 1,473

Text Arena · Creative Writing · No Style Control

AR · Chat / text · Human

It tests whether the model is actually useful in normal conversational turns, not just on narrow correctness tasks.

Rank #43 · Source label: gpt-5.4-high

verified runtimeexact alias

Raw row drilldownsource row, percentile, last updated, eligibility

Source: Arena
Raw value: 1,439
Percentile: 88.8%
Last updated: recent
Eligibility: headline eligible

Parsed from Arena leaderboard dataset row `gpt-5.4-high`. Category: creative_writing. Source rank: #45. Votes: 11109. Organization: openai. License: Proprietary.

88.8% percentile inside its fair comparison set

1,439Raw benchmark valueCI 1,432 - 1,446

Text Arena · English · No Style Control

AR · Chat / text · Human

It tests whether the model is actually useful in normal conversational turns, not just on narrow correctness tasks.

Rank #39 · Source label: gpt-5.4-high

verified runtimeexact alias

Raw row drilldownsource row, percentile, last updated, eligibility

Source: Arena
Raw value: 1,467
Percentile: 89.9%
Last updated: recent
Eligibility: headline eligible

Parsed from Arena leaderboard dataset row `gpt-5.4-high`. Category: english. Source rank: #41. Votes: 29222. Organization: openai. License: Proprietary.

89.9% percentile inside its fair comparison set

1,467Raw benchmark valueCI 1,462 - 1,472

Text Arena · Exclude Ties · No Style Control

AR · Chat / text · Human

It tests whether the model is actually useful in normal conversational turns, not just on narrow correctness tasks.

Rank #31 · Source label: gpt-5.4-high

verified runtimeexact alias

Raw row drilldownsource row, percentile, last updated, eligibility

Source: Arena
Raw value: 1,473
Percentile: 92%
Last updated: recent
Eligibility: headline eligible

Parsed from Arena leaderboard dataset row `gpt-5.4-high`. Category: exclude_ties. Source rank: #32. Votes: 49582. Organization: openai. License: Proprietary.

92% percentile inside its fair comparison set

1,473Raw benchmark valueCI 1,469 - 1,478

Text Arena · Hard Prompts · No Style Control

AR · Chat / text · Human

It tests whether the model is actually useful in normal conversational turns, not just on narrow correctness tasks.

Rank #26 · Source label: gpt-5.4-high

verified runtimeexact alias

Raw row drilldownsource row, percentile, last updated, eligibility

Source: Arena
Raw value: 1,485
Percentile: 93.4%
Last updated: recent
Eligibility: headline eligible

Parsed from Arena leaderboard dataset row `gpt-5.4-high`. Category: hard_prompts. Source rank: #27. Votes: 42068. Organization: openai. License: Proprietary.

93.4% percentile inside its fair comparison set

1,485Raw benchmark valueCI 1,481 - 1,490

Text Arena · Hard Prompts English · No Style Control

AR · Chat / text · Human

It tests whether the model is actually useful in normal conversational turns, not just on narrow correctness tasks.

Rank #36 · Source label: gpt-5.4-high

verified runtimeexact alias

Raw row drilldownsource row, percentile, last updated, eligibility

Source: Arena
Raw value: 1,479
Percentile: 90.6%
Last updated: recent
Eligibility: headline eligible

Parsed from Arena leaderboard dataset row `gpt-5.4-high`. Category: hard_prompts_english. Source rank: #37. Votes: 19665. Organization: openai. License: Proprietary.

90.6% percentile inside its fair comparison set

1,479Raw benchmark valueCI 1,473 - 1,484

Text Arena · Instruction Following · No Style Control

AR · Chat / text · Human

It tests whether the model is actually useful in normal conversational turns, not just on narrow correctness tasks.

Rank #28 · Source label: gpt-5.4-high

verified runtimeexact alias

Raw row drilldownsource row, percentile, last updated, eligibility

Source: Arena
Raw value: 1,469
Percentile: 92.8%
Last updated: recent
Eligibility: headline eligible

Parsed from Arena leaderboard dataset row `gpt-5.4-high`. Category: instruction_following. Source rank: #29. Votes: 21838. Organization: openai. License: Proprietary.

92.8% percentile inside its fair comparison set

1,469Raw benchmark valueCI 1,464 - 1,475

Text Arena · Longer Query · No Style Control

AR · Chat / text · Human

It tests whether the model is actually useful in normal conversational turns, not just on narrow correctness tasks.

Rank #38 · Source label: gpt-5.4-high

verified runtimeexact alias

Raw row drilldownsource row, percentile, last updated, eligibility

Source: Arena
Raw value: 1,474
Percentile: 89.6%
Last updated: recent
Eligibility: headline eligible

Parsed from Arena leaderboard dataset row `gpt-5.4-high`. Category: longer_query. Source rank: #40. Votes: 28402. Organization: openai. License: Proprietary.

89.6% percentile inside its fair comparison set

1,474Raw benchmark valueCI 1,468 - 1,479

Text Arena · Multi Turn · No Style Control

AR · Chat / text · Human

It tests whether the model is actually useful in normal conversational turns, not just on narrow correctness tasks.

Rank #20 · Source label: gpt-5.4-high

verified runtimeexact alias

Raw row drilldownsource row, percentile, last updated, eligibility

Source: Arena
Raw value: 1,482
Percentile: 94.9%
Last updated: recent
Eligibility: headline eligible

Parsed from Arena leaderboard dataset row `gpt-5.4-high`. Category: multi_turn. Source rank: #20. Votes: 11988. Organization: openai. License: Proprietary.

94.9% percentile inside its fair comparison set

1,482Raw benchmark valueCI 1,475 - 1,489

Instruction following

LB · Chat / text · Objective

It tests whether the model is actually useful in normal conversational turns, not just on narrow correctness tasks.

Rank #29 · Source label: gpt-5.4-xhigh

verified runtimeexact alias

Raw row drilldownsource row, percentile, last updated, eligibility

Source: LiveBench
Raw value: 70.2%
Percentile: 56.9%
Last updated: stale
Eligibility: headline eligible

Derived from the official LiveBench website leaderboard table. Category: IF. Tasks scored: 4.

56.9% percentile inside its fair comparison set

70.2%Raw benchmark value

Language

LB · Chat / text · Objective

It tests whether the model is actually useful in normal conversational turns, not just on narrow correctness tasks.

Rank #25 · Source label: gpt-5.4-xhigh

verified runtimeexact alias

Raw row drilldownsource row, percentile, last updated, eligibility

Source: LiveBench
Raw value: 82.6%
Percentile: 63.1%
Last updated: stale
Eligibility: headline eligible

Derived from the official LiveBench website leaderboard table. Category: Language. Tasks scored: 3.

63.1% percentile inside its fair comparison set

82.6%Raw benchmark value

Paraphrase

LB · Chat / text · Objective

It tests whether the model is actually useful in normal conversational turns, not just on narrow correctness tasks.

Rank #45 · Source label: gpt-5.4-xhigh

verified runtimeexact alias

Raw row drilldownsource row, percentile, last updated, eligibility

Source: LiveBench
Raw value: 63.4%
Percentile: 32.3%
Last updated: stale
Eligibility: headline eligible

Derived from the official LiveBench website leaderboard table. Task: paraphrase. Category: IF.

32.3% percentile inside its fair comparison set

63.4%Raw benchmark value

Simplify

LB · Chat / text · Objective

It tests whether the model is actually useful in normal conversational turns, not just on narrow correctness tasks.

Rank #15 · Source label: gpt-5.4-xhigh

verified runtimeexact alias

Raw row drilldownsource row, percentile, last updated, eligibility

Source: LiveBench
Raw value: 70%
Percentile: 78.5%
Last updated: stale
Eligibility: headline eligible

Derived from the official LiveBench website leaderboard table. Task: simplify. Category: IF.

78.5% percentile inside its fair comparison set

70%Raw benchmark value

Story generation

LB · Chat / text · Objective

It tests whether the model is actually useful in normal conversational turns, not just on narrow correctness tasks.

Rank #5 · Source label: gpt-5.4-xhigh

verified runtimeexact alias

Raw row drilldownsource row, percentile, last updated, eligibility

Source: LiveBench
Raw value: 79%
Percentile: 93.8%
Last updated: stale
Eligibility: headline eligible

Derived from the official LiveBench website leaderboard table. Task: story_generation. Category: IF.

93.8% percentile inside its fair comparison set

79%Raw benchmark value

Summarize

LB · Chat / text · Objective

It tests whether the model is actually useful in normal conversational turns, not just on narrow correctness tasks.

Rank #42 · Source label: gpt-5.4-xhigh

verified runtimeexact alias

Raw row drilldownsource row, percentile, last updated, eligibility

Source: LiveBench
Raw value: 68.4%
Percentile: 36.9%
Last updated: stale
Eligibility: headline eligible

Derived from the official LiveBench website leaderboard table. Task: summarize. Category: IF.

36.9% percentile inside its fair comparison set

68.4%Raw benchmark value

Connections

LB · Chat / text · Objective

It tests whether the model is actually useful in normal conversational turns, not just on narrow correctness tasks.

Rank #3 · Source label: gpt-5.4-xhigh

verified runtimeexact alias

Raw row drilldownsource row, percentile, last updated, eligibility

Source: LiveBench
Raw value: 100%
Percentile: 100%
Last updated: stale
Eligibility: headline eligible

Derived from the official LiveBench website leaderboard table. Task: connections. Category: Language.

100% percentile inside its fair comparison set

100%Raw benchmark value

Plot unscrambling

LB · Chat / text · Objective

It tests whether the model is actually useful in normal conversational turns, not just on narrow correctness tasks.

Rank #26 · Source label: gpt-5.4-xhigh

verified runtimeexact alias

Raw row drilldownsource row, percentile, last updated, eligibility

Source: LiveBench
Raw value: 65.9%
Percentile: 61.5%
Last updated: stale
Eligibility: headline eligible

Derived from the official LiveBench website leaderboard table. Task: plot_unscrambling. Category: Language.

61.5% percentile inside its fair comparison set

65.9%Raw benchmark value

Typos

LB · Chat / text · Objective

It tests whether the model is actually useful in normal conversational turns, not just on narrow correctness tasks.

Rank #28 · Source label: gpt-5.4-xhigh

verified runtimeexact alias

Raw row drilldownsource row, percentile, last updated, eligibility

Source: LiveBench
Raw value: 82%
Percentile: 61.5%
Last updated: stale
Eligibility: headline eligible

Derived from the official LiveBench website leaderboard table. Task: typos. Category: Language.

61.5% percentile inside its fair comparison set

82%Raw benchmark value

Coding31 benchmarks62.7%

Terminal-Bench Hard

AA · Coding · Objective

It tells you whether the model can generate, repair, and reason over code under evaluator pressure rather than marketing examples.

Rank #31 · Source label: GPT-5.4 (Non-reasoning)

verified runtimeexact alias

Raw row drilldownsource row, percentile, last updated, eligibility

Source: Artificial Analysis
Raw value: 37.9%
Percentile: 91%
Last updated: recent
Eligibility: headline eligible

Parsed from Artificial Analysis public leaderboard field `terminalbenchHard`.

91% percentile inside its fair comparison set

37.9%Raw benchmark value

HiL-Bench

SL · Coding · Rubric

It tells you whether the model can generate, repair, and reason over code under evaluator pressure rather than marketing examples.

Rank #15 · Source label: GPT-5.4

verified runtimeexact alias

Raw row drilldownsource row, percentile, last updated, eligibility

Source: Scale Labs
Raw value: 9.7%
Percentile: 12.5%
Last updated: recent
Eligibility: headline eligible

Parsed from the public Scale Labs page for scale-hil-bench. Reported model configuration: GPT-5.4. Collapse policy: highest reported score per canonical model.

12.5% percentile inside its fair comparison set

9.7%Raw benchmark value

Text Arena · Coding

AR · Coding · Human

It tells you whether the model can generate, repair, and reason over code under evaluator pressure rather than marketing examples.

Rank #35 · Source label: gpt-5.4-high

verified runtimeexact alias

Raw row drilldownsource row, percentile, last updated, eligibility

Source: Arena
Raw value: 1,521
Percentile: 90.8%
Last updated: recent
Eligibility: headline eligible

Parsed from Arena leaderboard dataset row `gpt-5.4-high`. Category: coding. Source rank: #35. Votes: 17566. Organization: openai. License: Proprietary.

90.8% percentile inside its fair comparison set

1,521Raw benchmark valueCI 1,515 - 1,527

Text Arena · Coding · No Style Control

AR · Coding · Human

It tells you whether the model can generate, repair, and reason over code under evaluator pressure rather than marketing examples.

Rank #28 · Source label: gpt-5.4-high

verified runtimeexact alias

Raw row drilldownsource row, percentile, last updated, eligibility

Source: Arena
Raw value: 1,497
Percentile: 92.7%
Last updated: recent
Eligibility: headline eligible

Parsed from Arena leaderboard dataset row `gpt-5.4-high`. Category: coding. Source rank: #28. Votes: 17566. Organization: openai. License: Proprietary.

92.7% percentile inside its fair comparison set

1,497Raw benchmark valueCI 1,491 - 1,502

Vibe Code Bench v1.1

VALS-AI · Coding · Objective

It tells you whether the model can generate, repair, and reason over code under evaluator pressure rather than marketing examples.

Rank #35 · Source label: openai/gpt-5.4-2026-03-05

verified runtimeexact alias

Raw row drilldownsource row, percentile, last updated, eligibility

Source: Vals AI
Raw value: 67.4%
Percentile: 63.4%
Last updated: recent
Eligibility: headline eligible

Parsed from Vals AI BenchmarkView overall scores. Vals slug: vibe-code; provider: OpenAI.

63.4% percentile inside its fair comparison set

67.4%Raw benchmark valueCI 57.9% - 76.9%

WebDev Arena

AR · Coding · Human

It tells you whether the model can generate, repair, and reason over code under evaluator pressure rather than marketing examples.

Rank #61 · Source label: gpt-5.4-high (codex-harness)

verified runtimeexact alias

Raw row drilldownsource row, percentile, last updated, eligibility

Source: Arena
Raw value: 1,465
Percentile: 52.8%
Last updated: recent
Eligibility: headline eligible

Parsed from Arena leaderboard dataset row `gpt-5.4-high (codex-harness)`. Category: overall. Source rank: #63. Votes: 1351. Organization: openai. License: Proprietary.

52.8% percentile inside its fair comparison set

1,465Raw benchmark valueCI 1,446 - 1,484

Code Arena · Webdev Html

AR · Coding · Human

It tells you whether the model can generate, repair, and reason over code under evaluator pressure rather than marketing examples.

Rank #58 · Source label: gpt-5.4-medium (codex-harness)

verified runtimeexact alias

Raw row drilldownsource row, percentile, last updated, eligibility

Source: Arena
Raw value: 1,478
Percentile: 54.4%
Last updated: recent
Eligibility: headline eligible

Parsed from Arena leaderboard dataset row `gpt-5.4-medium (codex-harness)`. Category: webdev-html. Source rank: #61. Votes: 132. Organization: openai. License: Proprietary.

54.4% percentile inside its fair comparison set

1,478Raw benchmark valueCI 1,425 - 1,531

Code Arena · Webdev React

AR · Coding · Human

It tells you whether the model can generate, repair, and reason over code under evaluator pressure rather than marketing examples.

Rank #61 · Source label: gpt-5.4-high (codex-harness)

verified runtimeexact alias

Raw row drilldownsource row, percentile, last updated, eligibility

Source: Arena
Raw value: 1,455
Percentile: 46.9%
Last updated: recent
Eligibility: headline eligible

Parsed from Arena leaderboard dataset row `gpt-5.4-high (codex-harness)`. Category: webdev-react. Source rank: #63. Votes: 1235. Organization: openai. License: Proprietary.

46.9% percentile inside its fair comparison set

1,455Raw benchmark valueCI 1,435 - 1,475

Code Migration

VALS-AI · Coding · Objective

It tells you whether the model can generate, repair, and reason over code under evaluator pressure rather than marketing examples.

Rank #33 · Source label: openai/gpt-5.4-2026-03-05

verified runtimeexact alias

Raw row drilldownsource row, percentile, last updated, eligibility

Source: Vals AI
Raw value: 35%
Percentile: 52.9%
Last updated: recent
Eligibility: headline eligible

Parsed from Vals AI BenchmarkView overall scores. Vals slug: code-migration; provider: OpenAI.

52.9% percentile inside its fair comparison set

35%Raw benchmark valueCI 26.9% - 43%

ProgramBench

VALS-AI · Coding · Objective

It tells you whether the model can generate, repair, and reason over code under evaluator pressure rather than marketing examples.

Rank #25 · Source label: openai/gpt-5.4-2026-03-05-high

verified runtimeexact alias

Raw row drilldownsource row, percentile, last updated, eligibility

Source: Vals AI
Raw value: 0.5%
Percentile: 72.4%
Last updated: recent
Eligibility: headline eligible

Parsed from Vals AI BenchmarkView overall scores. Vals slug: programbench; provider: OpenAI.

72.4% percentile inside its fair comparison set

0.5%Raw benchmark valueCI 0% - 1.5%

Code Arena · Image To Webdev

AR · Coding · Human

It tells you whether the model can generate, repair, and reason over code under evaluator pressure rather than marketing examples.

Rank #35 · Source label: gpt-5.4

verified runtimeexact alias

Raw row drilldownsource row, percentile, last updated, eligibility

Source: Arena
Raw value: 1,445
Percentile: 22.7%
Last updated: recent
Eligibility: headline eligible

Parsed from Arena leaderboard dataset row `gpt-5.4`. Category: image_to_webdev. Source rank: #38. Votes: 2433. Organization: openai. License: Proprietary.

22.7% percentile inside its fair comparison set

1,445Raw benchmark valueCI 1,432 - 1,459

SWE-bench Verified

VALS-AI · Coding · Objective

It tells you whether the model can generate, repair, and reason over code under evaluator pressure rather than marketing examples.

Rank #28 · Source label: openai/gpt-5.4-2026-03-05

verified runtimeexact alias

Raw row drilldownsource row, percentile, last updated, eligibility

Source: Vals AI
Raw value: 78.2%
Percentile: 67.1%
Last updated: recent
Eligibility: headline eligible

Parsed from Vals AI BenchmarkView overall scores. Vals slug: swebench; provider: OpenAI.

67.1% percentile inside its fair comparison set

78.2%Raw benchmark valueCI 74.6% - 81.8%

LiveCodeBench

VALS-AI · Coding · Objective

It tells you whether the model can generate, repair, and reason over code under evaluator pressure rather than marketing examples.

Rank #39 · Source label: openai/gpt-5.4-2026-03-05

verified runtimeexact alias

Raw row drilldownsource row, percentile, last updated, eligibility

Source: Vals AI
Raw value: 84.1%
Percentile: 69.6%
Last updated: recent
Eligibility: headline eligible

Parsed from Vals AI BenchmarkView overall scores. Vals slug: lcb; provider: OpenAI.

69.6% percentile inside its fair comparison set

84.1%Raw benchmark valueCI 82.1% - 86.2%

Agentic coding

LB · Coding · Objective

It tells you whether the model can generate, repair, and reason over code under evaluator pressure rather than marketing examples.

Rank #34 · Source label: gpt-5.4-xhigh

verified runtimeexact alias

Raw row drilldownsource row, percentile, last updated, eligibility

Source: LiveBench
Raw value: 53.8%
Percentile: 49.2%
Last updated: stale
Eligibility: headline eligible

Derived from the official LiveBench website leaderboard table. Category: Agentic Coding. Tasks scored: 3.

49.2% percentile inside its fair comparison set

53.8%Raw benchmark value

Coding

LB · Coding · Objective

It tells you whether the model can generate, repair, and reason over code under evaluator pressure rather than marketing examples.

Rank #36 · Source label: gpt-5.4-xhigh

verified runtimeexact alias

Raw row drilldownsource row, percentile, last updated, eligibility

Source: LiveBench
Raw value: 77.5%
Percentile: 46.2%
Last updated: stale
Eligibility: headline eligible

Derived from the official LiveBench website leaderboard table. Category: Coding. Tasks scored: 2.

46.2% percentile inside its fair comparison set

77.5%Raw benchmark value

JavaScript

LB · Coding · Objective

It tells you whether the model can generate, repair, and reason over code under evaluator pressure rather than marketing examples.

Rank #17 · Source label: gpt-5.4-xhigh

verified runtimeexact alias

Raw row drilldownsource row, percentile, last updated, eligibility

Source: LiveBench
Raw value: 68.2%
Percentile: 78.5%
Last updated: stale
Eligibility: headline eligible

Derived from the official LiveBench website leaderboard table. Task: javascript. Category: Agentic Coding.

78.5% percentile inside its fair comparison set

68.2%Raw benchmark value

TypeScript

LB · Coding · Objective

It tells you whether the model can generate, repair, and reason over code under evaluator pressure rather than marketing examples.

Rank #27 · Source label: gpt-5.4-xhigh

verified runtimeexact alias

Raw row drilldownsource row, percentile, last updated, eligibility

Source: LiveBench
Raw value: 43.3%
Percentile: 61.5%
Last updated: stale
Eligibility: headline eligible

Derived from the official LiveBench website leaderboard table. Task: typescript. Category: Agentic Coding.

61.5% percentile inside its fair comparison set

43.3%Raw benchmark value

Python

LB · Coding · Objective

It tells you whether the model can generate, repair, and reason over code under evaluator pressure rather than marketing examples.

Rank #42 · Source label: gpt-5.4-xhigh

verified runtimeexact alias

Raw row drilldownsource row, percentile, last updated, eligibility

Source: LiveBench
Raw value: 50%
Percentile: 46.2%
Last updated: stale
Eligibility: headline eligible

Derived from the official LiveBench website leaderboard table. Task: python. Category: Agentic Coding.

46.2% percentile inside its fair comparison set

50%Raw benchmark value

Coding generation

LB · Coding · Objective

It tells you whether the model can generate, repair, and reason over code under evaluator pressure rather than marketing examples.

Rank #47 · Source label: gpt-5.4-xhigh

verified runtimeexact alias

Raw row drilldownsource row, percentile, last updated, eligibility

Source: LiveBench
Raw value: 74.6%
Percentile: 30.8%
Last updated: stale
Eligibility: headline eligible

Derived from the official LiveBench website leaderboard table. Task: code_generation. Category: Coding.

30.8% percentile inside its fair comparison set

74.6%Raw benchmark value

Coding completion

LB · Coding · Objective

It tells you whether the model can generate, repair, and reason over code under evaluator pressure rather than marketing examples.

Rank #17 · Source label: gpt-5.4-xhigh

verified runtimeexact alias

Raw row drilldownsource row, percentile, last updated, eligibility

Source: LiveBench
Raw value: 80.4%
Percentile: 78.5%
Last updated: stale
Eligibility: headline eligible

Derived from the official LiveBench website leaderboard table. Task: code_completion. Category: Coding.

78.5% percentile inside its fair comparison set

80.4%Raw benchmark value

Debugging

BB · Coding · Rubric

It tells you whether the model can generate, repair, and reason over code under evaluator pressure rather than marketing examples.

Rank #9 · Source label: openai/gpt-5.4

verified runtimeexact alias

Raw row drilldownsource row, percentile, last updated, eligibility

Source: BridgeBench
Raw value: 85.6%
Percentile: 66.7%
Last updated: archived
Eligibility: headline eligible

Parsed from the BridgeBench page for bridgebench-debugging.

66.7% percentile inside its fair comparison set

85.6%Raw benchmark value

Refactoring

BB · Coding · Rubric

It tells you whether the model can generate, repair, and reason over code under evaluator pressure rather than marketing examples.

Rank #12 · Source label: openai/gpt-5.4

verified runtimeexact alias

Raw row drilldownsource row, percentile, last updated, eligibility

Source: BridgeBench
Raw value: 63.4%
Percentile: 47.6%
Last updated: archived
Eligibility: headline eligible

Parsed from the BridgeBench page for bridgebench-refactoring.

47.6% percentile inside its fair comparison set

63.4%Raw benchmark value

Security

BB · Coding · Rubric

It tells you whether the model can generate, repair, and reason over code under evaluator pressure rather than marketing examples.

Rank #4 · Source label: openai/gpt-5.4

verified runtimeexact alias

Raw row drilldownsource row, percentile, last updated, eligibility

Source: BridgeBench
Raw value: 84.4%
Percentile: 88%
Last updated: archived
Eligibility: headline eligible

Parsed from the BridgeBench page for bridgebench-security.

88% percentile inside its fair comparison set

84.4%Raw benchmark value

UI bench

BB · Coding · Rubric

It tells you whether the model can generate, repair, and reason over code under evaluator pressure rather than marketing examples.

Rank #11 · Source label: openai/gpt-5.4

verified runtimeexact alias

Raw row drilldownsource row, percentile, last updated, eligibility

Source: BridgeBench
Raw value: 65.7%
Percentile: 54.5%
Last updated: archived
Eligibility: headline eligible

Parsed from the BridgeBench page for bridgebench-ui.

54.5% percentile inside its fair comparison set

65.7%Raw benchmark value

Speed throughput

BB · Coding · Speed / cost

It tells you whether the model can generate, repair, and reason over code under evaluator pressure rather than marketing examples.

Rank #18 · Source label: openai/gpt-5.4

verified runtimeexact alias

Raw row drilldownsource row, percentile, last updated, eligibility

Source: BridgeBench
Raw value: 88 t/s
Percentile: 32%
Last updated: archived
Eligibility: headline eligible

Parsed from the BridgeBench page for bridgebench-throughput.

32% percentile inside its fair comparison set

88 t/sRaw benchmark value

Speed TTFT

BB · Coding · Speed / cost

It tells you whether the model can generate, repair, and reason over code under evaluator pressure rather than marketing examples.

Rank #2 · Source label: openai/gpt-5.4

verified runtimeexact alias

Raw row drilldownsource row, percentile, last updated, eligibility

Source: BridgeBench
Raw value: 397.00ms
Percentile: 96%
Last updated: archived
Eligibility: headline eligible

Parsed from the BridgeBench page for bridgebench-ttft.

96% percentile inside its fair comparison set

397.00msRaw benchmark value

Speed cost

BB · Coding · Speed / cost

It tells you whether the model can generate, repair, and reason over code under evaluator pressure rather than marketing examples.

Rank #21 · Source label: openai/gpt-5.4

verified runtimeexact alias

Raw row drilldownsource row, percentile, last updated, eligibility

Source: BridgeBench
Raw value: $0.3 $
Percentile: 20%
Last updated: archived
Eligibility: headline eligible

Parsed from the BridgeBench page for bridgebench-speed-cost.

20% percentile inside its fair comparison set

$0.3 $Raw benchmark value

Terminal-Bench 2.0

OFF · Coding · Objective

It tells you whether the model can generate, repair, and reason over code under evaluator pressure rather than marketing examples.

Rank #2

Official company resultmanual verifiedmanual verified

Raw row drilldownsource row, percentile, last updated, eligibility

Source: Provider official evals
Raw value: 75.1%
Percentile: 83.3%
Last updated: archived
Eligibility: headline eligible

Verified against OpenAI's GPT-5.4 launch page dated March 5, 2026. Uses the published Terminal-Bench 2.0 figure. Checked 2026-04-29. Verification: manual_public_page_verification.

83.3% percentile inside its fair comparison set

75.1%Raw benchmark value

Overall quality

BB · Coding · Combined

It tells you whether the model can generate, repair, and reason over code under evaluator pressure rather than marketing examples.

Rank #8 · Source label: GPT-5.4

verified runtimeexact alias

Raw row drilldownsource row, percentile, last updated, eligibility

Source: BridgeBench
Raw value: 86.3%
Percentile: 68.2%
Last updated: archived
Eligibility: headline eligible

Parsed from the BridgeBench page for bridgebench-overall-quality.

68.2% percentile inside its fair comparison set

86.3%Raw benchmark value

Overall vibe

BB · Coding · Combined

It tells you whether the model can generate, repair, and reason over code under evaluator pressure rather than marketing examples.

Rank #9 · Source label: GPT-5.4

verified runtimeexact alias

Raw row drilldownsource row, percentile, last updated, eligibility

Source: BridgeBench
Raw value: 79.6%
Percentile: 63.6%
Last updated: archived
Eligibility: headline eligible

Parsed from the BridgeBench page for bridgebench-overall-vibe.

63.6% percentile inside its fair comparison set

79.6%Raw benchmark value

Overall coverage

BB · Coding · Combined

It tells you whether the model can generate, repair, and reason over code under evaluator pressure rather than marketing examples.

Rank #7 · Source label: GPT-5.4

verified runtimeexact alias

Raw row drilldownsource row, percentile, last updated, eligibility

Source: BridgeBench
Raw value: 100%
Percentile: 100%
Last updated: archived
Eligibility: headline eligible

Parsed from the BridgeBench page for bridgebench-overall-coverage.

100% percentile inside its fair comparison set

100%Raw benchmark value

Reasoning / math / science21 benchmarks72.7%

Humanity's Last Exam

AA · Reasoning / math / science · Objective

It is one of the cleaner reads on deliberate reasoning strength rather than style or popularity.

Rank #210 · Source label: GPT-5.4 (Non-reasoning)

verified runtimeexact alias

Raw row drilldownsource row, percentile, last updated, eligibility

Source: Artificial Analysis
Raw value: 11.3%
Percentile: 60.5%
Last updated: recent
Eligibility: headline eligible

Parsed from Artificial Analysis public leaderboard field `hle`.

60.5% percentile inside its fair comparison set

11.3%Raw benchmark value

GPQA

AA · Reasoning / math / science · Objective

It is one of the cleaner reads on deliberate reasoning strength rather than style or popularity.

Rank #183 · Source label: GPT-5.4 (Non-reasoning)

verified runtimeexact alias

Raw row drilldownsource row, percentile, last updated, eligibility

Source: Artificial Analysis
Raw value: 74.8%
Percentile: 63.1%
Last updated: recent
Eligibility: headline eligible

Parsed from Artificial Analysis public leaderboard field `gpqa`.

63.1% percentile inside its fair comparison set

74.8%Raw benchmark value

CritPt

AA · Reasoning / math / science · Objective

It is one of the cleaner reads on deliberate reasoning strength rather than style or popularity.

Rank #171 · Source label: GPT-5.4 (Non-reasoning)

verified runtimeexact alias

Raw row drilldownsource row, percentile, last updated, eligibility

Source: Artificial Analysis
Raw value: 0.6%
Percentile: 62.8%
Last updated: recent
Eligibility: headline eligible

Parsed from Artificial Analysis public leaderboard field `critpt`.

62.8% percentile inside its fair comparison set

0.6%Raw benchmark value

TutorBench

SL · Reasoning / math / science · Rubric

It is one of the cleaner reads on deliberate reasoning strength rather than style or popularity.

Rank #2 · Source label: gpt-5.4-pro-2026-03-05

verified runtimeexact alias

Raw row drilldownsource row, percentile, last updated, eligibility

Source: Scale Labs
Raw value: 56.6%
Percentile: 96.3%
Last updated: recent
Eligibility: headline eligible

Parsed from the public Scale Labs page for scale-tutorbench. Reported model configuration: gpt-5.4-pro-2026-03-05. Collapse policy: highest reported score per canonical model.

96.3% percentile inside its fair comparison set

56.6%Raw benchmark value

MultiNRC

SL · Reasoning / math / science · Rubric

It is one of the cleaner reads on deliberate reasoning strength rather than style or popularity.

Rank #6 · Source label: gpt-5.4-pro-2026-03-05

verified runtimeexact alias

Raw row drilldownsource row, percentile, last updated, eligibility

Source: Scale Labs
Raw value: 62.3%
Percentile: 88.6%
Last updated: recent
Eligibility: headline eligible

Parsed from the public Scale Labs page for scale-multinrc. Reported model configuration: gpt-5.4-pro-2026-03-05. Collapse policy: highest reported score per canonical model.

88.6% percentile inside its fair comparison set

62.3%Raw benchmark value

Text Arena · Math

AR · Reasoning / math / science · Human

It is one of the cleaner reads on deliberate reasoning strength rather than style or popularity.

Rank #24 · Source label: gpt-5.4-high

verified runtimeexact alias

Raw row drilldownsource row, percentile, last updated, eligibility

Source: Arena
Raw value: 1,492
Percentile: 93.6%
Last updated: recent
Eligibility: headline eligible

Parsed from Arena leaderboard dataset row `gpt-5.4-high`. Category: math. Source rank: #25. Votes: 3379. Organization: openai. License: Proprietary.

93.6% percentile inside its fair comparison set

1,492Raw benchmark valueCI 1,481 - 1,502

Text Arena · Math · No Style Control

AR · Reasoning / math / science · Human

It is one of the cleaner reads on deliberate reasoning strength rather than style or popularity.

Rank #22 · Source label: gpt-5.4-high

verified runtimeexact alias

Raw row drilldownsource row, percentile, last updated, eligibility

Source: Arena
Raw value: 1,488
Percentile: 94.2%
Last updated: recent
Eligibility: headline eligible

Parsed from Arena leaderboard dataset row `gpt-5.4-high`. Category: math. Source rank: #22. Votes: 3379. Organization: openai. License: Proprietary.

94.2% percentile inside its fair comparison set

1,488Raw benchmark valueCI 1,477 - 1,498

GPQA Diamond

VALS-AI · Reasoning / math / science · Objective

It is one of the cleaner reads on deliberate reasoning strength rather than style or popularity.

Rank #21 · Source label: openai/gpt-5.4-2026-03-05

verified runtimeexact alias

Raw row drilldownsource row, percentile, last updated, eligibility

Source: Vals AI
Raw value: 91.7%
Percentile: 86.3%
Last updated: recent
Eligibility: headline eligible

Parsed from Vals AI BenchmarkView overall scores. Vals slug: gpqa; provider: OpenAI.

86.3% percentile inside its fair comparison set

91.7%Raw benchmark valueCI 87.9% - 95.4%

MMLU Pro

VALS-AI · Reasoning / math / science · Objective

It is one of the cleaner reads on deliberate reasoning strength rather than style or popularity.

Rank #27 · Source label: openai/gpt-5.4-2026-03-05

verified runtimeexact alias

Raw row drilldownsource row, percentile, last updated, eligibility

Source: Vals AI
Raw value: 87.5%
Percentile: 79.2%
Last updated: recent
Eligibility: headline eligible

Parsed from Vals AI BenchmarkView overall scores. Vals slug: mmlu_pro; provider: OpenAI.

79.2% percentile inside its fair comparison set

87.5%Raw benchmark valueCI 86.7% - 88.3%

Mathematics

LB · Reasoning / math / science · Objective

It is one of the cleaner reads on deliberate reasoning strength rather than style or popularity.

Rank #19 · Source label: gpt-5.4-xhigh

verified runtimeexact alias

Raw row drilldownsource row, percentile, last updated, eligibility

Source: LiveBench
Raw value: 94.1%
Percentile: 72.3%
Last updated: stale
Eligibility: headline eligible

Derived from the official LiveBench website leaderboard table. Category: Mathematics. Tasks scored: 4.

72.3% percentile inside its fair comparison set

94.1%Raw benchmark value

Reasoning

LB · Reasoning / math / science · Objective

It is one of the cleaner reads on deliberate reasoning strength rather than style or popularity.

Rank #24 · Source label: gpt-5.4-xhigh

verified runtimeexact alias

Raw row drilldownsource row, percentile, last updated, eligibility

Source: LiveBench
Raw value: 88.1%
Percentile: 64.6%
Last updated: stale
Eligibility: headline eligible

Derived from the official LiveBench website leaderboard table. Category: Reasoning. Tasks scored: 4.

64.6% percentile inside its fair comparison set

88.1%Raw benchmark value

AMPS Hard

LB · Reasoning / math / science · Objective

It is one of the cleaner reads on deliberate reasoning strength rather than style or popularity.

Rank #21 · Source label: gpt-5.4-xhigh

verified runtimeexact alias

Raw row drilldownsource row, percentile, last updated, eligibility

Source: LiveBench
Raw value: 98%
Percentile: 83.1%
Last updated: stale
Eligibility: headline eligible

Derived from the official LiveBench website leaderboard table. Task: AMPS_Hard. Category: Mathematics.

83.1% percentile inside its fair comparison set

98%Raw benchmark value

Integrals with game

LB · Reasoning / math / science · Objective

It is one of the cleaner reads on deliberate reasoning strength rather than style or popularity.

Rank #18 · Source label: gpt-5.4-xhigh

verified runtimeexact alias

Raw row drilldownsource row, percentile, last updated, eligibility

Source: LiveBench
Raw value: 93%
Percentile: 73.8%
Last updated: stale
Eligibility: headline eligible

Derived from the official LiveBench website leaderboard table. Task: integrals_with_game. Category: Mathematics.

73.8% percentile inside its fair comparison set

93%Raw benchmark value

Math competition

LB · Reasoning / math / science · Objective

It is one of the cleaner reads on deliberate reasoning strength rather than style or popularity.

Rank #51 · Source label: gpt-5.4-xhigh

verified runtimeexact alias

Raw row drilldownsource row, percentile, last updated, eligibility

Source: LiveBench
Raw value: 94.1%
Percentile: 26.2%
Last updated: stale
Eligibility: headline eligible

Derived from the official LiveBench website leaderboard table. Task: math_comp. Category: Mathematics.

26.2% percentile inside its fair comparison set

94.1%Raw benchmark value

Olympiad

LB · Reasoning / math / science · Objective

It is one of the cleaner reads on deliberate reasoning strength rather than style or popularity.

Rank #23 · Source label: gpt-5.4-xhigh

verified runtimeexact alias

Raw row drilldownsource row, percentile, last updated, eligibility

Source: LiveBench
Raw value: 91.5%
Percentile: 66.2%
Last updated: stale
Eligibility: headline eligible

Derived from the official LiveBench website leaderboard table. Task: olympiad. Category: Mathematics.

66.2% percentile inside its fair comparison set

91.5%Raw benchmark value

Theory of mind

LB · Reasoning / math / science · Objective

It is one of the cleaner reads on deliberate reasoning strength rather than style or popularity.

Rank #1 · Source label: gpt-5.4-xhigh

verified runtimeexact alias

Raw row drilldownsource row, percentile, last updated, eligibility

Source: LiveBench
Raw value: 88.5%
Percentile: 100%
Last updated: stale
Eligibility: headline eligible

Derived from the official LiveBench website leaderboard table. Task: theory_of_mind. Category: Reasoning.

100% percentile inside its fair comparison set

88.5%Raw benchmark value

Zebra puzzle

LB · Reasoning / math / science · Objective

It is one of the cleaner reads on deliberate reasoning strength rather than style or popularity.

Rank #25 · Source label: gpt-5.4-xhigh

verified runtimeexact alias

Raw row drilldownsource row, percentile, last updated, eligibility

Source: LiveBench
Raw value: 98%
Percentile: 64.6%
Last updated: stale
Eligibility: headline eligible

Derived from the official LiveBench website leaderboard table. Task: zebra_puzzle. Category: Reasoning.

64.6% percentile inside its fair comparison set

98%Raw benchmark value

Spatial

LB · Reasoning / math / science · Objective

It is one of the cleaner reads on deliberate reasoning strength rather than style or popularity.

Rank #21 · Source label: gpt-5.4-xhigh

verified runtimeexact alias

Raw row drilldownsource row, percentile, last updated, eligibility

Source: LiveBench
Raw value: 98%
Percentile: 72.3%
Last updated: stale
Eligibility: headline eligible

Derived from the official LiveBench website leaderboard table. Task: spatial. Category: Reasoning.

72.3% percentile inside its fair comparison set

98%Raw benchmark value

Logic with navigation

LB · Reasoning / math / science · Objective

It is one of the cleaner reads on deliberate reasoning strength rather than style or popularity.

Rank #48 · Source label: gpt-5.4-xhigh

verified runtimeexact alias

Raw row drilldownsource row, percentile, last updated, eligibility

Source: LiveBench
Raw value: 68%
Percentile: 30.8%
Last updated: stale
Eligibility: headline eligible

Derived from the official LiveBench website leaderboard table. Task: logic_with_navigation. Category: Reasoning.

30.8% percentile inside its fair comparison set

68%Raw benchmark value

Reasoning

BB · Reasoning / math / science · Rubric

It is one of the cleaner reads on deliberate reasoning strength rather than style or popularity.

Rank #3 · Source label: openai/gpt-5.4

verified runtimeexact alias

Raw row drilldownsource row, percentile, last updated, eligibility

Source: BridgeBench
Raw value: 40.6%
Percentile: 90.5%
Last updated: archived
Eligibility: headline eligible

Parsed from the BridgeBench page for bridgebench-reasoning.

90.5% percentile inside its fair comparison set

40.6%Raw benchmark value

Humanity's Last Exam

OFF · Reasoning / math / science · Objective

It is one of the cleaner reads on deliberate reasoning strength rather than style or popularity.

Rank #5

Official company resultmanual verifiedmanual verified

Raw row drilldownsource row, percentile, last updated, eligibility

Source: Provider official evals
Raw value: 39.8%
Percentile: 42.9%
Last updated: archived
Eligibility: headline eligible

Verified against OpenAI's GPT-5.4 launch page dated March 5, 2026. Uses the published no-tools Humanity's Last Exam figure. Checked 2026-04-29. Verification: manual_public_page_verification.

42.9% percentile inside its fair comparison set

39.8%Raw benchmark value

Professional reasoning35 benchmarks66.6%

APEX-Agents-AA

AA · Professional reasoning · Objective

Long-horizon agentic task completion.

Rank #9 · Source label: GPT-5.4 (Xhigh)

verified runtimeexact alias

Raw row drilldownsource row, percentile, last updated, eligibility

Source: Artificial Analysis
Raw value: 33.3%
Percentile: 75%
Last updated: recent
Eligibility: headline eligible

Parsed from Artificial Analysis public leaderboard field `apexAgents`.

75% percentile inside its fair comparison set

33.3%Raw benchmark value

PRBench Legal

SL · Professional reasoning · Rubric

Applied legal reasoning on professional-domain tasks.

Rank #19 · Source label: gpt-5.4 (High)

verified runtimeexact alias

Raw row drilldownsource row, percentile, last updated, eligibility

Source: Scale Labs
Raw value: 44.4%
Percentile: 56.1%
Last updated: recent
Eligibility: headline eligible

Parsed from the public Scale Labs page for scale-prbench-legal. Reported model configuration: gpt-5.4 (High). Collapse policy: highest reported score per canonical model.

56.1% percentile inside its fair comparison set

44.4%Raw benchmark value

Text Arena · Expert

AR · Professional reasoning · Human

Observed user preference in Arena's Text Arena expert leaderboard.

Rank #23 · Source label: gpt-5.4-high

verified runtimeexact alias

Raw row drilldownsource row, percentile, last updated, eligibility

Source: Arena
Raw value: 1,515
Percentile: 93.3%
Last updated: recent
Eligibility: headline eligible

Parsed from Arena leaderboard dataset row `gpt-5.4-high`. Category: expert. Source rank: #23. Votes: 6333. Organization: openai. License: Proprietary.

93.3% percentile inside its fair comparison set

1,515Raw benchmark valueCI 1,507 - 1,524

Text Arena · Industry Business And Management And Financial Operations

AR · Professional reasoning · Human

Observed user preference in Arena's Text Arena industry_business_and_management_and_financial_operations leaderboard.

Rank #24 · Source label: gpt-5.4-high

verified runtimeexact alias

Raw row drilldownsource row, percentile, last updated, eligibility

Source: Arena
Raw value: 1,482
Percentile: 93.8%
Last updated: recent
Eligibility: headline eligible

93.8% percentile inside its fair comparison set

1,482Raw benchmark valueCI 1,475 - 1,488

Text Arena · Industry Entertainment And Sports And Media

AR · Professional reasoning · Human

Observed user preference in Arena's Text Arena industry_entertainment_and_sports_and_media leaderboard.

Rank #47 · Source label: gpt-5.4-high

verified runtimeexact alias

Raw row drilldownsource row, percentile, last updated, eligibility

Source: Arena
Raw value: 1,441
Percentile: 87.7%
Last updated: recent
Eligibility: headline eligible

Parsed from Arena leaderboard dataset row `gpt-5.4-high`. Category: industry_entertainment_and_sports_and_media. Source rank: #50. Votes: 14124. Organization: openai. License: Proprietary.

87.7% percentile inside its fair comparison set

1,441Raw benchmark valueCI 1,435 - 1,448

Text Arena · Industry Legal And Government

AR · Professional reasoning · Human

Observed user preference in Arena's Text Arena industry_legal_and_government leaderboard.

Rank #26 · Source label: gpt-5.4-high

verified runtimeexact alias

Raw row drilldownsource row, percentile, last updated, eligibility

Source: Arena
Raw value: 1,490
Percentile: 92.8%
Last updated: recent
Eligibility: headline eligible

Parsed from Arena leaderboard dataset row `gpt-5.4-high`. Category: industry_legal_and_government. Source rank: #26. Votes: 5228. Organization: openai. License: Proprietary.

92.8% percentile inside its fair comparison set

1,490Raw benchmark valueCI 1,481 - 1,499

Text Arena · Industry Life And Physical And Social Science

AR · Professional reasoning · Human

Observed user preference in Arena's Text Arena industry_life_and_physical_and_social_science leaderboard.

Rank #42 · Source label: gpt-5.4-high

verified runtimeexact alias

Raw row drilldownsource row, percentile, last updated, eligibility

Source: Arena
Raw value: 1,488
Percentile: 89%
Last updated: recent
Eligibility: headline eligible

Parsed from Arena leaderboard dataset row `gpt-5.4-high`. Category: industry_life_and_physical_and_social_science. Source rank: #44. Votes: 10640. Organization: openai. License: Proprietary.

89% percentile inside its fair comparison set

1,488Raw benchmark valueCI 1,481 - 1,495

Text Arena · Industry Mathematical

AR · Professional reasoning · Human

Observed user preference in Arena's Text Arena industry_mathematical leaderboard.

Rank #19 · Source label: gpt-5.4-high

verified runtimeexact alias

Raw row drilldownsource row, percentile, last updated, eligibility

Source: Arena
Raw value: 1,499
Percentile: 94.9%
Last updated: recent
Eligibility: headline eligible

Parsed from Arena leaderboard dataset row `gpt-5.4-high`. Category: industry_mathematical. Source rank: #19. Votes: 3543. Organization: openai. License: Proprietary.

94.9% percentile inside its fair comparison set

1,499Raw benchmark valueCI 1,488 - 1,510

Text Arena · Industry Medicine And Healthcare

AR · Professional reasoning · Human

Observed user preference in Arena's Text Arena industry_medicine_and_healthcare leaderboard.

Rank #60 · Source label: gpt-5.4-high

verified runtimeexact alias

Raw row drilldownsource row, percentile, last updated, eligibility

Source: Arena
Raw value: 1,475
Percentile: 82.9%
Last updated: recent
Eligibility: headline eligible

Parsed from Arena leaderboard dataset row `gpt-5.4-high`. Category: industry_medicine_and_healthcare. Source rank: #66. Votes: 4838. Organization: openai. License: Proprietary.

82.9% percentile inside its fair comparison set

1,475Raw benchmark valueCI 1,465 - 1,485

Text Arena · Industry Software And It Services

AR · Professional reasoning · Human

Observed user preference in Arena's Text Arena industry_software_and_it_services leaderboard.

Rank #41 · Source label: gpt-5.4-high

verified runtimeexact alias

Raw row drilldownsource row, percentile, last updated, eligibility

Source: Arena
Raw value: 1,507
Percentile: 89.4%
Last updated: recent
Eligibility: headline eligible

Parsed from Arena leaderboard dataset row `gpt-5.4-high`. Category: industry_software_and_it_services. Source rank: #41. Votes: 25424. Organization: openai. License: Proprietary.

89.4% percentile inside its fair comparison set

1,507Raw benchmark valueCI 1,502 - 1,512

Text Arena · Industry Writing And Literature And Language

AR · Professional reasoning · Human

Observed user preference in Arena's Text Arena industry_writing_and_literature_and_language leaderboard.

Rank #30 · Source label: gpt-5.4-high

verified runtimeexact alias

Raw row drilldownsource row, percentile, last updated, eligibility

Source: Arena
Raw value: 1,465
Percentile: 92.3%
Last updated: recent
Eligibility: headline eligible

Parsed from Arena leaderboard dataset row `gpt-5.4-high`. Category: industry_writing_and_literature_and_language. Source rank: #32. Votes: 16324. Organization: openai. License: Proprietary.

92.3% percentile inside its fair comparison set

1,465Raw benchmark valueCI 1,459 - 1,471

Text Arena · Expert · No Style Control

AR · Professional reasoning · Human

Observed user preference in Arena's Text Arena expert leaderboard.

Rank #22 · Source label: gpt-5.4-high

verified runtimeexact alias

Raw row drilldownsource row, percentile, last updated, eligibility

Source: Arena
Raw value: 1,507
Percentile: 93.6%
Last updated: recent
Eligibility: headline eligible

Parsed from Arena leaderboard dataset row `gpt-5.4-high`. Category: expert. Source rank: #22. Votes: 6333. Organization: openai. License: Proprietary.

93.6% percentile inside its fair comparison set

1,507Raw benchmark valueCI 1,499 - 1,515

Text Arena · Industry Business And Management And Financial Operations · No Style Control

AR · Professional reasoning · Human

Observed user preference in Arena's Text Arena industry_business_and_management_and_financial_operations leaderboard.

Rank #17 · Source label: gpt-5.4-high

verified runtimeexact alias

Raw row drilldownsource row, percentile, last updated, eligibility

Source: Arena
Raw value: 1,475
Percentile: 95.7%
Last updated: recent
Eligibility: headline eligible

95.7% percentile inside its fair comparison set

1,475Raw benchmark valueCI 1,468 - 1,481

Text Arena · Industry Entertainment And Sports And Media · No Style Control

AR · Professional reasoning · Human

Observed user preference in Arena's Text Arena industry_entertainment_and_sports_and_media leaderboard.

Rank #38 · Source label: gpt-5.4-high

verified runtimeexact alias

Raw row drilldownsource row, percentile, last updated, eligibility

Source: Arena
Raw value: 1,438
Percentile: 90.1%
Last updated: recent
Eligibility: headline eligible

Parsed from Arena leaderboard dataset row `gpt-5.4-high`. Category: industry_entertainment_and_sports_and_media. Source rank: #40. Votes: 14124. Organization: openai. License: Proprietary.

90.1% percentile inside its fair comparison set

1,438Raw benchmark valueCI 1,432 - 1,444

Text Arena · Industry Legal And Government · No Style Control

AR · Professional reasoning · Human

Observed user preference in Arena's Text Arena industry_legal_and_government leaderboard.

Rank #21 · Source label: gpt-5.4-high

verified runtimeexact alias

Raw row drilldownsource row, percentile, last updated, eligibility

Source: Arena
Raw value: 1,486
Percentile: 94.3%
Last updated: recent
Eligibility: headline eligible

Parsed from Arena leaderboard dataset row `gpt-5.4-high`. Category: industry_legal_and_government. Source rank: #22. Votes: 5228. Organization: openai. License: Proprietary.

94.3% percentile inside its fair comparison set

1,486Raw benchmark valueCI 1,477 - 1,495

Text Arena · Industry Life And Physical And Social Science · No Style Control

AR · Professional reasoning · Human

Observed user preference in Arena's Text Arena industry_life_and_physical_and_social_science leaderboard.

Rank #32 · Source label: gpt-5.4-high

verified runtimeexact alias

Raw row drilldownsource row, percentile, last updated, eligibility

Source: Arena
Raw value: 1,478
Percentile: 91.7%
Last updated: recent
Eligibility: headline eligible

Parsed from Arena leaderboard dataset row `gpt-5.4-high`. Category: industry_life_and_physical_and_social_science. Source rank: #34. Votes: 10640. Organization: openai. License: Proprietary.

91.7% percentile inside its fair comparison set

1,478Raw benchmark valueCI 1,472 - 1,485

Text Arena · Industry Mathematical · No Style Control

AR · Professional reasoning · Human

Observed user preference in Arena's Text Arena industry_mathematical leaderboard.

Rank #16 · Source label: gpt-5.4-high

verified runtimeexact alias

Raw row drilldownsource row, percentile, last updated, eligibility

Source: Arena
Raw value: 1,497
Percentile: 95.8%
Last updated: recent
Eligibility: headline eligible

Parsed from Arena leaderboard dataset row `gpt-5.4-high`. Category: industry_mathematical. Source rank: #16. Votes: 3543. Organization: openai. License: Proprietary.

95.8% percentile inside its fair comparison set

1,497Raw benchmark valueCI 1,486 - 1,507

Text Arena · Industry Medicine And Healthcare · No Style Control

AR · Professional reasoning · Human

Observed user preference in Arena's Text Arena industry_medicine_and_healthcare leaderboard.

Rank #52 · Source label: gpt-5.4-high

verified runtimeexact alias

Raw row drilldownsource row, percentile, last updated, eligibility

Source: Arena
Raw value: 1,459
Percentile: 85.3%
Last updated: recent
Eligibility: headline eligible

Parsed from Arena leaderboard dataset row `gpt-5.4-high`. Category: industry_medicine_and_healthcare. Source rank: #53. Votes: 4838. Organization: openai. License: Proprietary.

85.3% percentile inside its fair comparison set

1,459Raw benchmark valueCI 1,449 - 1,468

Text Arena · Industry Software And It Services · No Style Control

AR · Professional reasoning · Human

Observed user preference in Arena's Text Arena industry_software_and_it_services leaderboard.

Rank #30 · Source label: gpt-5.4-high

verified runtimeexact alias

Raw row drilldownsource row, percentile, last updated, eligibility

Source: Arena
Raw value: 1,491
Percentile: 92.3%
Last updated: recent
Eligibility: headline eligible

Parsed from Arena leaderboard dataset row `gpt-5.4-high`. Category: industry_software_and_it_services. Source rank: #30. Votes: 25424. Organization: openai. License: Proprietary.

92.3% percentile inside its fair comparison set

1,491Raw benchmark valueCI 1,486 - 1,496

Text Arena · Industry Writing And Literature And Language · No Style Control

AR · Professional reasoning · Human

Observed user preference in Arena's Text Arena industry_writing_and_literature_and_language leaderboard.

Rank #31 · Source label: gpt-5.4-high

verified runtimeexact alias

Raw row drilldownsource row, percentile, last updated, eligibility

Source: Arena
Raw value: 1,458
Percentile: 92%
Last updated: recent
Eligibility: headline eligible

Parsed from Arena leaderboard dataset row `gpt-5.4-high`. Category: industry_writing_and_literature_and_language. Source rank: #33. Votes: 16324. Organization: openai. License: Proprietary.

92% percentile inside its fair comparison set

1,458Raw benchmark valueCI 1,453 - 1,464

Harvey's Legal Agent Benchmark

VALS-AI · Professional reasoning · Objective

Completing legal work with documents, spreadsheets, presentations, and file-system tools.

Rank #68 · Source label: openai/gpt-5.4-2026-03-05

verified runtimeexact alias

Raw row drilldownsource row, percentile, last updated, eligibility

Source: Vals AI
Raw value: 0%
Percentile: 20.3%
Last updated: recent
Eligibility: headline eligible

Parsed from Vals AI BenchmarkView overall scores. Vals slug: hlab; provider: OpenAI.

20.3% percentile inside its fair comparison set

0%Raw benchmark valueCI 0% - 0%

LegalBench

VALS-AI · Professional reasoning · Objective

Academic legal reasoning tasks.

Rank #15 · Source label: openai/gpt-5.4-2026-03-05

verified runtimeexact alias

Raw row drilldownsource row, percentile, last updated, eligibility

Source: Vals AI
Raw value: 86%
Percentile: 89.6%
Last updated: recent
Eligibility: headline eligible

Parsed from Vals AI BenchmarkView overall scores. Vals slug: legal_bench; provider: OpenAI.

89.6% percentile inside its fair comparison set

86%Raw benchmark valueCI 85.2% - 86.9%

MedCode

VALS-AI · Professional reasoning · Objective

Medical billing support and coding tasks.

Rank #54 · Source label: openai/gpt-5.4-2026-03-05

verified runtimeexact alias

Raw row drilldownsource row, percentile, last updated, eligibility

Source: Vals AI
Raw value: 41.3%
Percentile: 44.2%
Last updated: recent
Eligibility: headline eligible

Parsed from Vals AI BenchmarkView overall scores. Vals slug: medcode; provider: OpenAI.

44.2% percentile inside its fair comparison set

41.3%Raw benchmark valueCI 37.1% - 45.5%

MedScribe

VALS-AI · Professional reasoning · Objective

Administrative documentation support for doctors.

Rank #62 · Source label: openai/gpt-5.4-2026-03-05

verified runtimeexact alias

Raw row drilldownsource row, percentile, last updated, eligibility

Source: Vals AI
Raw value: 77.5%
Percentile: 36.5%
Last updated: recent
Eligibility: headline eligible

Parsed from Vals AI BenchmarkView overall scores. Vals slug: medscribe; provider: OpenAI.

36.5% percentile inside its fair comparison set

77.5%Raw benchmark valueCI 71% - 84%

SAGE

VALS-AI · Professional reasoning · Objective

Student Assessment with Generative Evaluation.

Rank #48 · Source label: openai/gpt-5.4-2026-03-05

verified runtimeexact alias

Raw row drilldownsource row, percentile, last updated, eligibility

Source: Vals AI
Raw value: 43.3%
Percentile: 42%
Last updated: recent
Eligibility: headline eligible

Parsed from Vals AI BenchmarkView overall scores. Vals slug: sage; provider: OpenAI.

42% percentile inside its fair comparison set

43.3%Raw benchmark valueCI 37.2% - 49.4%

SkillsBench

VALS-AI · Professional reasoning · Objective

Applied professional skills tasks.

Rank #18 · Source label: openai/gpt-5.4-2026-03-05

verified runtimeexact alias

Raw row drilldownsource row, percentile, last updated, eligibility

Source: Vals AI
Raw value: 51.7%
Percentile: 48.5%
Last updated: recent
Eligibility: headline eligible

Parsed from Vals AI BenchmarkView overall scores. Vals slug: skillsbench; provider: OpenAI.

48.5% percentile inside its fair comparison set

51.7%Raw benchmark valueCI 42.5% - 61%

TaxEval v2

VALS-AI · Professional reasoning · Objective

Answer quality on tax questions and responses.

Rank #44 · Source label: openai/gpt-5.4-2026-03-05

verified runtimeexact alias

Raw row drilldownsource row, percentile, last updated, eligibility

Source: Vals AI
Raw value: 74%
Percentile: 67.4%
Last updated: recent
Eligibility: headline eligible

Parsed from Vals AI BenchmarkView overall scores. Vals slug: tax_eval_v2; provider: OpenAI.

67.4% percentile inside its fair comparison set

74%Raw benchmark valueCI 72.3% - 75.7%

Data analysis

LB · Professional reasoning · Objective

Structured data manipulation and table reasoning accuracy.

Rank #15 · Source label: gpt-5.4-xhigh

verified runtimeexact alias

Raw row drilldownsource row, percentile, last updated, eligibility

Source: LiveBench
Raw value: 79.3%
Percentile: 78.5%
Last updated: stale
Eligibility: headline eligible

Derived from the official LiveBench website leaderboard table. Category: Data Analysis. Tasks scored: 3.

78.5% percentile inside its fair comparison set

79.3%Raw benchmark value

Overall

LB · Professional reasoning · Objective

Average objective performance across LiveBench's current public category mix.

Rank #19 · Source label: gpt-5.4-xhigh

verified runtimeexact alias

Raw row drilldownsource row, percentile, last updated, eligibility

Source: LiveBench
Raw value: 78%
Percentile: 72.3%
Last updated: stale
Eligibility: headline eligible

Derived from the official LiveBench website leaderboard table. Category averages included: 7.

72.3% percentile inside its fair comparison set

78%Raw benchmark value

Consecutive events

LB · Professional reasoning · Objective

Objective consecutive events score in LiveBench.

Rank #28 · Source label: gpt-5.4-xhigh

verified runtimeexact alias

Raw row drilldownsource row, percentile, last updated, eligibility

Source: LiveBench
Raw value: 86.2%
Percentile: 58.5%
Last updated: stale
Eligibility: headline eligible

Derived from the official LiveBench website leaderboard table. Task: consecutive_events. Category: Data Analysis.

58.5% percentile inside its fair comparison set

86.2%Raw benchmark value

Table join

LB · Professional reasoning · Objective

Objective table join score in LiveBench.

Rank #12 · Source label: gpt-5.4-xhigh

verified runtimeexact alias

Raw row drilldownsource row, percentile, last updated, eligibility

Source: LiveBench
Raw value: 51.8%
Percentile: 83.1%
Last updated: stale
Eligibility: headline eligible

Derived from the official LiveBench website leaderboard table. Task: tablejoin. Category: Data Analysis.

83.1% percentile inside its fair comparison set

51.8%Raw benchmark value

Table reformat

LB · Professional reasoning · Objective

Objective table reformat score in LiveBench.

Rank #5 · Source label: gpt-5.4-xhigh

verified runtimeexact alias

Raw row drilldownsource row, percentile, last updated, eligibility

Source: LiveBench
Raw value: 100%
Percentile: 100%
Last updated: stale
Eligibility: headline eligible

Derived from the official LiveBench website leaderboard table. Task: tablereformat. Category: Data Analysis.

100% percentile inside its fair comparison set

100%Raw benchmark value

Hallucination

BB · Professional reasoning · Rubric

Accuracy and fabrication resistance under prompts that invite unsupported claims.

Rank #15 · Source label: openai/gpt-5.4

verified runtimeexact alias

Raw row drilldownsource row, percentile, last updated, eligibility

Source: BridgeBench
Raw value: 72.8%
Percentile: 58.8%
Last updated: archived
Eligibility: headline eligible

Parsed from the BridgeBench page for bridgebench-hallucination.

58.8% percentile inside its fair comparison set

72.8%Raw benchmark value

BS pushback

BB · Professional reasoning · Rubric

Resistance to confidently accepting bogus assumptions in expert-style prompts.

Rank #4 · Source label: openai/gpt-5.4

verified runtimeexact alias

Raw row drilldownsource row, percentile, last updated, eligibility

Source: BridgeBench
Raw value: 91.5%
Percentile: 88.9%
Last updated: archived
Eligibility: headline eligible

Parsed from the BridgeBench page for bridgebench-pushback.

88.9% percentile inside its fair comparison set

91.5%Raw benchmark value

Poker Agent

VALS-AI · Professional reasoning · Objective

Agent profit in poker-style strategic play.

Rank #3 · Source label: openai/gpt-5-2025-08-07

backfilledproxy backfilledBackground only

Raw row drilldownsource row, percentile, last updated, eligibility

Source: Vals AI
Raw value: 1,103.2 score
Percentile: 94.7%
Last updated: archived
Eligibility: Fallback benchmark identity is visible for context but excluded from default ranking.

Parsed from Vals AI BenchmarkView overall scores. Vals slug: poker_agent; provider: unknown. Backfilled from GPT-5 via approved benchmark identity mapping map-gpt-5-4-to-gpt-5.

94.7% percentile inside its fair comparison set

1,103.2 scoreRaw benchmark valueCI 1,103.2 score - 1,103.2 score

Search / tool use4 benchmarks54.6%

Tau2-Bench Telecom

AA · Search / tool use · Objective

It matters when the model must browse, call tools, and recover useful answers from external systems.

Rank #170 · Source label: GPT-5.4 (Non-reasoning)

verified runtimeexact alias

Raw row drilldownsource row, percentile, last updated, eligibility

Source: Artificial Analysis
Raw value: 36%
Percentile: 50.1%
Last updated: recent
Eligibility: headline eligible

Parsed from Artificial Analysis public leaderboard field `tau2`.

50.1% percentile inside its fair comparison set

36%Raw benchmark value

Search Arena

AR · Search / tool use · Human

It matters when the model must browse, call tools, and recover useful answers from external systems.

Rank #13 · Source label: gpt-5.4-search

verified runtimeexact alias

Raw row drilldownsource row, percentile, last updated, eligibility

Source: Arena
Raw value: 1,195
Percentile: 63.6%
Last updated: recent
Eligibility: headline eligible

Parsed from Arena leaderboard dataset row `gpt-5.4-search`. Category: overall. Source rank: #14. Votes: 110116. Organization: openai. License: Proprietary.

63.6% percentile inside its fair comparison set

1,195Raw benchmark valueCI 1,189 - 1,200

Search Arena · No Style Control

AR · Search / tool use · Human

It matters when the model must browse, call tools, and recover useful answers from external systems.

Rank #16 · Source label: gpt-5.4-search

verified runtimeexact alias

Raw row drilldownsource row, percentile, last updated, eligibility

Source: Arena
Raw value: 1,197
Percentile: 54.5%
Last updated: recent
Eligibility: headline eligible

Parsed from Arena leaderboard dataset row `gpt-5.4-search`. Category: overall. Source rank: #16. Votes: 110116. Organization: openai. License: Proprietary.

54.5% percentile inside its fair comparison set

1,197Raw benchmark valueCI 1,192 - 1,203

BrowseComp

OFF · Search / tool use · Objective

It matters when the model must browse, call tools, and recover useful answers from external systems.

Rank #4

Official company resultmanual verifiedmanual verified

Raw row drilldownsource row, percentile, last updated, eligibility

Source: Provider official evals
Raw value: 82.7%
Percentile: 50%
Last updated: archived
Eligibility: headline eligible

Verified against OpenAI's GPT-5.4 launch page dated March 5, 2026. Uses the published BrowseComp figure. Checked 2026-04-29. Verification: manual_public_page_verification.

50% percentile inside its fair comparison set

82.7%Raw benchmark value

Long context2 benchmarks64%

Long Context Reasoning

AA · Long context · Objective

It checks whether long-context claims survive contact with retrieval, memory, or long-document tasks.

Rank #204 · Source label: GPT-5.4 (Non-reasoning)

verified runtimeexact alias

Raw row drilldownsource row, percentile, last updated, eligibility

Source: Artificial Analysis
Raw value: 58.3%
Percentile: 55.4%
Last updated: recent
Eligibility: headline eligible

Parsed from Artificial Analysis public leaderboard field `lcr`.

55.4% percentile inside its fair comparison set

58.3%Raw benchmark value

CorpFin v2

VALS-AI · Long context · Objective

It checks whether long-context claims survive contact with retrieval, memory, or long-document tasks.

Rank #33 · Source label: openai/gpt-5.4-2026-03-05

verified runtimeexact alias

Raw row drilldownsource row, percentile, last updated, eligibility

Source: Vals AI
Raw value: 65.3%
Percentile: 72.6%
Last updated: aging
Eligibility: headline eligible

Parsed from Vals AI BenchmarkView overall scores. Vals slug: corp_fin_v2; provider: OpenAI.

72.6% percentile inside its fair comparison set

65.3%Raw benchmark valueCI 63.4% - 67.1%

Vision understanding25 benchmarks84.8%

MMMU-Pro

AA · Vision understanding · Objective

It is useful when the model must read charts, UI, screenshots, or visual scenes rather than text alone.

Rank #89 · Source label: GPT-5.4 (Non-reasoning)

verified runtimeexact alias

Raw row drilldownsource row, percentile, last updated, eligibility

Source: Artificial Analysis
Raw value: 70.6%
Percentile: 57.7%
Last updated: recent
Eligibility: headline eligible

Parsed from Artificial Analysis public leaderboard field `mmmuPro`.

57.7% percentile inside its fair comparison set

70.6%Raw benchmark value

VISTA

SL · Vision understanding · Rubric

It is useful when the model must read charts, UI, screenshots, or visual scenes rather than text alone.

Rank #3 · Source label: gpt-5.4-pro-2026-03-05

verified runtimeexact alias

Raw row drilldownsource row, percentile, last updated, eligibility

Source: Scale Labs
Raw value: 53.9%
Percentile: 96.8%
Last updated: recent
Eligibility: headline eligible

Parsed from the public Scale Labs page for scale-vista. Reported model configuration: gpt-5.4-pro-2026-03-05. Collapse policy: highest reported score per canonical model.

96.8% percentile inside its fair comparison set

53.9%Raw benchmark value

VTB

SL · Vision understanding · Rubric

It is useful when the model must read charts, UI, screenshots, or visual scenes rather than text alone.

Rank #2 · Source label: gpt-5.4-2026-03-05 (reasoning effort = high)

verified runtimeexact alias

Raw row drilldownsource row, percentile, last updated, eligibility

Source: Scale Labs
Raw value: 29.2%
Percentile: 95.8%
Last updated: recent
Eligibility: headline eligible

Parsed from the public Scale Labs page for scale-vtb. Reported model configuration: gpt-5.4-2026-03-05 (reasoning effort = high). Collapse policy: highest reported score per canonical model.

95.8% percentile inside its fair comparison set

29.2%Raw benchmark value

Vision Arena

AR · Vision understanding · Human

It is useful when the model must read charts, UI, screenshots, or visual scenes rather than text alone.

Rank #21 · Source label: gpt-5.4-high

verified runtimeexact alias

Raw row drilldownsource row, percentile, last updated, eligibility

Source: Arena
Raw value: 1,284
Percentile: 85.6%
Last updated: recent
Eligibility: headline eligible

Parsed from Arena leaderboard dataset row `gpt-5.4-high`. Category: overall. Source rank: #22. Votes: 29177. Organization: openai. License: Proprietary.

85.6% percentile inside its fair comparison set

1,284Raw benchmark valueCI 1,278 - 1,290

Vision Arena · Captioning

AR · Vision understanding · Human

It is useful when the model must read charts, UI, screenshots, or visual scenes rather than text alone.

Rank #15 · Source label: gpt-5-chat

backfilledproxy backfilledBackground only

Raw row drilldownsource row, percentile, last updated, eligibility

Source: Arena
Raw value: 1,195
Percentile: 58.1%
Last updated: recent
Eligibility: Fallback benchmark identity is visible for context but excluded from default ranking.

58.1% percentile inside its fair comparison set

1,195Raw benchmark valueCI 1,163 - 1,227

Vision Arena · Creative Writing Vision

AR · Vision understanding · Human

It is useful when the model must read charts, UI, screenshots, or visual scenes rather than text alone.

Rank #27 · Source label: gpt-5.4-high

verified runtimeexact alias

Raw row drilldownsource row, percentile, last updated, eligibility

Source: Arena
Raw value: 1,260
Percentile: 68.3%
Last updated: recent
Eligibility: headline eligible

Parsed from Arena leaderboard dataset row `gpt-5.4-high`. Category: creative_writing_vision. Source rank: #31. Votes: 1945. Organization: openai. License: Proprietary.

68.3% percentile inside its fair comparison set

1,260Raw benchmark valueCI 1,245 - 1,274

Vision Arena · Diagram

AR · Vision understanding · Human

It is useful when the model must read charts, UI, screenshots, or visual scenes rather than text alone.

Rank #13 · Source label: gpt-5.4-high

verified runtimeexact alias

Raw row drilldownsource row, percentile, last updated, eligibility

Source: Arena
Raw value: 1,321
Percentile: 87.9%
Last updated: recent
Eligibility: headline eligible

Parsed from Arena leaderboard dataset row `gpt-5.4-high`. Category: diagram. Source rank: #13. Votes: 7803. Organization: openai. License: Proprietary.

87.9% percentile inside its fair comparison set

1,321Raw benchmark valueCI 1,312 - 1,329

Vision Arena · English

AR · Vision understanding · Human

It is useful when the model must read charts, UI, screenshots, or visual scenes rather than text alone.

Rank #16 · Source label: gpt-5.4-high

verified runtimeexact alias

Raw row drilldownsource row, percentile, last updated, eligibility

Source: Arena
Raw value: 1,282
Percentile: 89.1%
Last updated: recent
Eligibility: headline eligible

Parsed from Arena leaderboard dataset row `gpt-5.4-high`. Category: english. Source rank: #16. Votes: 11572. Organization: openai. License: Proprietary.

89.1% percentile inside its fair comparison set

1,282Raw benchmark valueCI 1,274 - 1,290

Vision Arena · Entity Recognition

AR · Vision understanding · Human

It is useful when the model must read charts, UI, screenshots, or visual scenes rather than text alone.

Rank #21 · Source label: gpt-5.4

verified runtimeexact alias

Raw row drilldownsource row, percentile, last updated, eligibility

Source: Arena
Raw value: 1,215
Percentile: 50%
Last updated: recent
Eligibility: headline eligible

Parsed from Arena leaderboard dataset row `gpt-5.4`. Category: entity_recognition. Source rank: #23. Votes: 134. Organization: openai. License: Proprietary.

50% percentile inside its fair comparison set

1,215Raw benchmark valueCI 1,168 - 1,262

Vision Arena · Homework

AR · Vision understanding · Human

It is useful when the model must read charts, UI, screenshots, or visual scenes rather than text alone.

Rank #9 · Source label: gpt-5.4-high

verified runtimeexact alias

Raw row drilldownsource row, percentile, last updated, eligibility

Source: Arena
Raw value: 1,331
Percentile: 91.7%
Last updated: recent
Eligibility: headline eligible

Parsed from Arena leaderboard dataset row `gpt-5.4-high`. Category: homework. Source rank: #10. Votes: 3768. Organization: openai. License: Proprietary.

91.7% percentile inside its fair comparison set

1,331Raw benchmark valueCI 1,321 - 1,342

Vision Arena · Humor

AR · Vision understanding · Human

It is useful when the model must read charts, UI, screenshots, or visual scenes rather than text alone.

Rank #22 · Source label: gpt-5.4-high

verified runtimeexact alias

Raw row drilldownsource row, percentile, last updated, eligibility

Source: Arena
Raw value: 1,263
Percentile: 71.2%
Last updated: recent
Eligibility: headline eligible

Parsed from Arena leaderboard dataset row `gpt-5.4-high`. Category: humor. Source rank: #23. Votes: 1061. Organization: openai. License: Proprietary.

71.2% percentile inside its fair comparison set

1,263Raw benchmark valueCI 1,244 - 1,283

Vision Arena · Ocr

AR · Vision understanding · Human

It is useful when the model must read charts, UI, screenshots, or visual scenes rather than text alone.

Rank #19 · Source label: gpt-5.4-high

verified runtimeexact alias

Raw row drilldownsource row, percentile, last updated, eligibility

Source: Arena
Raw value: 1,299
Percentile: 82%
Last updated: recent
Eligibility: headline eligible

Parsed from Arena leaderboard dataset row `gpt-5.4-high`. Category: ocr. Source rank: #19. Votes: 20737. Organization: openai. License: Proprietary.

82% percentile inside its fair comparison set

1,299Raw benchmark valueCI 1,293 - 1,304

Vision Arena · No Style Control

AR · Vision understanding · Human

It is useful when the model must read charts, UI, screenshots, or visual scenes rather than text alone.

Rank #16 · Source label: gpt-5.4-high

verified runtimeexact alias

Raw row drilldownsource row, percentile, last updated, eligibility

Source: Arena
Raw value: 1,302
Percentile: 89.2%
Last updated: recent
Eligibility: headline eligible

Parsed from Arena leaderboard dataset row `gpt-5.4-high`. Category: overall. Source rank: #17. Votes: 29177. Organization: openai. License: Proprietary.

89.2% percentile inside its fair comparison set

1,302Raw benchmark valueCI 1,296 - 1,308

Vision Arena · Captioning · No Style Control

AR · Vision understanding · Human

It is useful when the model must read charts, UI, screenshots, or visual scenes rather than text alone.

Rank #19 · Source label: gpt-5-chat

backfilledproxy backfilledBackground only

Raw row drilldownsource row, percentile, last updated, eligibility

Source: Arena
Raw value: 1,212
Percentile: 45.2%
Last updated: recent
Eligibility: Fallback benchmark identity is visible for context but excluded from default ranking.

45.2% percentile inside its fair comparison set

1,212Raw benchmark valueCI 1,181 - 1,244

Vision Arena · Creative Writing Vision · No Style Control

AR · Vision understanding · Human

It is useful when the model must read charts, UI, screenshots, or visual scenes rather than text alone.

Rank #23 · Source label: gpt-5.4-high

verified runtimeexact alias

Raw row drilldownsource row, percentile, last updated, eligibility

Source: Arena
Raw value: 1,281
Percentile: 73.2%
Last updated: recent
Eligibility: headline eligible

Parsed from Arena leaderboard dataset row `gpt-5.4-high`. Category: creative_writing_vision. Source rank: #26. Votes: 1945. Organization: openai. License: Proprietary.

73.2% percentile inside its fair comparison set

1,281Raw benchmark valueCI 1,267 - 1,295

Vision Arena · Diagram · No Style Control

AR · Vision understanding · Human

It is useful when the model must read charts, UI, screenshots, or visual scenes rather than text alone.

Rank #11 · Source label: gpt-5.4-high

verified runtimeexact alias

Raw row drilldownsource row, percentile, last updated, eligibility

Source: Arena
Raw value: 1,326
Percentile: 89.9%
Last updated: recent
Eligibility: headline eligible

Parsed from Arena leaderboard dataset row `gpt-5.4-high`. Category: diagram. Source rank: #11. Votes: 7803. Organization: openai. License: Proprietary.

89.9% percentile inside its fair comparison set

1,326Raw benchmark valueCI 1,318 - 1,335

Vision Arena · English · No Style Control

AR · Vision understanding · Human

It is useful when the model must read charts, UI, screenshots, or visual scenes rather than text alone.

Rank #13 · Source label: gpt-5.4-high

verified runtimeexact alias

Raw row drilldownsource row, percentile, last updated, eligibility

Source: Arena
Raw value: 1,301
Percentile: 91.3%
Last updated: recent
Eligibility: headline eligible

Parsed from Arena leaderboard dataset row `gpt-5.4-high`. Category: english. Source rank: #14. Votes: 11572. Organization: openai. License: Proprietary.

91.3% percentile inside its fair comparison set

1,301Raw benchmark valueCI 1,293 - 1,310

Vision Arena · Entity Recognition · No Style Control

AR · Vision understanding · Human

It is useful when the model must read charts, UI, screenshots, or visual scenes rather than text alone.

Rank #15 · Source label: gpt-5.4

verified runtimeexact alias

Raw row drilldownsource row, percentile, last updated, eligibility

Source: Arena
Raw value: 1,251
Percentile: 65%
Last updated: recent
Eligibility: headline eligible

Parsed from Arena leaderboard dataset row `gpt-5.4`. Category: entity_recognition. Source rank: #17. Votes: 134. Organization: openai. License: Proprietary.

65% percentile inside its fair comparison set

1,251Raw benchmark valueCI 1,205 - 1,298

Vision Arena · Homework · No Style Control

AR · Vision understanding · Human

It is useful when the model must read charts, UI, screenshots, or visual scenes rather than text alone.

Rank #11 · Source label: gpt-5.4-high

verified runtimeexact alias

Raw row drilldownsource row, percentile, last updated, eligibility

Source: Arena
Raw value: 1,337
Percentile: 89.6%
Last updated: recent
Eligibility: headline eligible

Parsed from Arena leaderboard dataset row `gpt-5.4-high`. Category: homework. Source rank: #11. Votes: 3768. Organization: openai. License: Proprietary.

89.6% percentile inside its fair comparison set

1,337Raw benchmark valueCI 1,327 - 1,348

Vision Arena · Humor · No Style Control

AR · Vision understanding · Human

It is useful when the model must read charts, UI, screenshots, or visual scenes rather than text alone.

Rank #15 · Source label: gpt-5.4-high

verified runtimeexact alias

Raw row drilldownsource row, percentile, last updated, eligibility

Source: Arena
Raw value: 1,284
Percentile: 80.8%
Last updated: recent
Eligibility: headline eligible

Parsed from Arena leaderboard dataset row `gpt-5.4-high`. Category: humor. Source rank: #16. Votes: 1061. Organization: openai. License: Proprietary.

80.8% percentile inside its fair comparison set

1,284Raw benchmark valueCI 1,265 - 1,303

Vision Arena · Ocr · No Style Control

AR · Vision understanding · Human

It is useful when the model must read charts, UI, screenshots, or visual scenes rather than text alone.

Rank #14 · Source label: gpt-5.4-high

verified runtimeexact alias

Raw row drilldownsource row, percentile, last updated, eligibility

Source: Arena
Raw value: 1,311
Percentile: 87%
Last updated: recent
Eligibility: headline eligible

Parsed from Arena leaderboard dataset row `gpt-5.4-high`. Category: ocr. Source rank: #15. Votes: 20737. Organization: openai. License: Proprietary.

87% percentile inside its fair comparison set

1,311Raw benchmark valueCI 1,305 - 1,317

MMMU Pro

VALS-AI · Vision understanding · Objective

It is useful when the model must read charts, UI, screenshots, or visual scenes rather than text alone.

Rank #15 · Source label: openai/gpt-5.4-2026-03-05

verified runtimeexact alias

Raw row drilldownsource row, percentile, last updated, eligibility

Source: Vals AI
Raw value: 87.5%
Percentile: 84.3%
Last updated: recent
Eligibility: headline eligible

Parsed from Vals AI BenchmarkView overall scores. Vals slug: mmmu; provider: OpenAI.

84.3% percentile inside its fair comparison set

87.5%Raw benchmark valueCI 86% - 89.1%

MMMU-Pro

OFF · Vision understanding · Objective

It is useful when the model must read charts, UI, screenshots, or visual scenes rather than text alone.

Rank #2

Official company resultmanual verifiedmanual verified

Raw row drilldownsource row, percentile, last updated, eligibility

Source: Provider official evals
Raw value: 81.2%
Percentile: 100%
Last updated: archived
Eligibility: headline eligible

Verified against OpenAI's GPT-5.4 launch page dated March 5, 2026. Uses the published no-tools MMMU Pro figure. Checked 2026-04-29. Verification: manual_public_page_verification.

100% percentile inside its fair comparison set

81.2%Raw benchmark value

Vision Arena · Creative Writing

AR · Vision understanding · Human

It is useful when the model must read charts, UI, screenshots, or visual scenes rather than text alone.

Rank #12 · Source label: gpt-5-chat

backfilledproxy backfilledBackground only

Raw row drilldownsource row, percentile, last updated, eligibility

Source: Arena
Raw value: 1,232
Percentile: 68.8%
Last updated: archived
Eligibility: Fallback benchmark identity is visible for context but excluded from default ranking.

68.8% percentile inside its fair comparison set

1,232Raw benchmark valueCI 1,216 - 1,248

Vision Arena · Creative Writing · No Style Control

AR · Vision understanding · Human

It is useful when the model must read charts, UI, screenshots, or visual scenes rather than text alone.

Rank #11 · Source label: gpt-5-chat

backfilledproxy backfilledBackground only

Raw row drilldownsource row, percentile, last updated, eligibility

Source: Arena
Raw value: 1,239
Percentile: 71.9%
Last updated: archived
Eligibility: Fallback benchmark identity is visible for context but excluded from default ranking.

71.9% percentile inside its fair comparison set

1,239Raw benchmark valueCI 1,224 - 1,255

Document understanding4 benchmarks70.7%

Document Arena

AR · Document understanding · Human

It matters when the job is reading PDFs, tables, forms, or mixed-layout documents rather than plain chat.

Rank #17 · Source label: gpt-5.4

verified runtimeexact alias

Raw row drilldownsource row, percentile, last updated, eligibility

Source: Arena
Raw value: 1,470
Percentile: 60%
Last updated: recent
Eligibility: headline eligible

Parsed from Arena leaderboard dataset row `gpt-5.4`. Category: overall. Source rank: #18. Votes: 33331. Organization: openai. License: Proprietary.

60% percentile inside its fair comparison set

1,470Raw benchmark valueCI 1,464 - 1,476

Document Arena · No Style Control

AR · Document understanding · Human

It matters when the job is reading PDFs, tables, forms, or mixed-layout documents rather than plain chat.

Rank #13 · Source label: gpt-5.4

verified runtimeexact alias

Raw row drilldownsource row, percentile, last updated, eligibility

Source: Arena
Raw value: 1,471
Percentile: 70%
Last updated: recent
Eligibility: headline eligible

Parsed from Arena leaderboard dataset row `gpt-5.4`. Category: overall. Source rank: #14. Votes: 33331. Organization: openai. License: Proprietary.

70% percentile inside its fair comparison set

1,471Raw benchmark valueCI 1,465 - 1,477

MortgageTax

VALS-AI · Document understanding · Objective

It matters when the job is reading PDFs, tables, forms, or mixed-layout documents rather than plain chat.

Rank #17 · Source label: openai/gpt-5.4-2026-03-05

verified runtimeexact alias

Raw row drilldownsource row, percentile, last updated, eligibility

Source: Vals AI
Raw value: 68.3%
Percentile: 81.4%
Last updated: recent
Eligibility: headline eligible

Parsed from Vals AI BenchmarkView overall scores. Vals slug: mortgage_tax; provider: OpenAI.

81.4% percentile inside its fair comparison set

68.3%Raw benchmark valueCI 66.5% - 70.1%

Multimodal mix

OC · Document understanding · Objective

It matters when the job is reading PDFs, tables, forms, or mixed-layout documents rather than plain chat.

Rank #3 · Source label: gpt-5

backfilledproxy backfilledBackground only

Raw row drilldownsource row, percentile, last updated, eligibility

Source: OpenCompass
Raw value: 75.4%
Percentile: 92.9%
Last updated: archived
Eligibility: Fallback benchmark identity is visible for context but excluded from default ranking.

Backfilled from GPT-5 via approved benchmark identity mapping map-gpt-5-4-to-gpt-5.

92.9% percentile inside its fair comparison set

75.4%Raw benchmark value

Safety1 benchmark91%

MASK

SL · Safety · Rubric

Whether a model stays honest instead of covertly optimizing against the user.

Rank #7 · Source label: gpt-5.4-pro-2026-03-05

verified runtimeexact alias

Raw row drilldownsource row, percentile, last updated, eligibility

Source: Scale Labs
Raw value: 91.7%
Percentile: 91%
Last updated: recent
Eligibility: headline eligible

Parsed from the public Scale Labs page for scale-mask. Reported model configuration: gpt-5.4-pro-2026-03-05. Collapse policy: highest reported score per canonical model.

91% percentile inside its fair comparison set

91.7%Raw benchmark value

Embeddings / retrieval1 benchmarkn/a

Retrieval

MTEB · Embeddings / retrieval · Retrieval

It is one of the few direct signals for retrieval stacks, where embedding quality matters more than chat style.

Rank #2 · Source label: gpt-5

backfilledproxy backfilledBackground only

Raw row drilldownsource row, percentile, last updated, eligibility

Source: MTEB
Raw value: 58.8 ndcg
Percentile: 100%
Last updated: archived
Eligibility: Fallback benchmark identity is visible for context but excluded from default ranking.

Embedding endpoint score. Backfilled from GPT-5 via approved benchmark identity mapping map-gpt-5-4-to-gpt-5.

100% percentile inside its fair comparison set

58.8 ndcgRaw benchmark value

Multilingual20 benchmarksn/a

Text Arena · Chinese

AR · Multilingual · Human

Observed user preference in Arena's Text Arena chinese leaderboard.

Rank #46 · Source label: gpt-5.4-high

verified runtimeexact alias

Raw row drilldownsource row, percentile, last updated, eligibility

Source: Arena
Raw value: 1,507
Percentile: 87.2%
Last updated: recent
Eligibility: headline eligible

Parsed from Arena leaderboard dataset row `gpt-5.4-high`. Category: chinese. Source rank: #48. Votes: 4132. Organization: openai. License: Proprietary.

87.2% percentile inside its fair comparison set

1,507Raw benchmark valueCI 1,497 - 1,517

Text Arena · French

AR · Multilingual · Human

Observed user preference in Arena's Text Arena french leaderboard.

Rank #24 · Source label: gpt-5.4-high

verified runtimeexact alias

Raw row drilldownsource row, percentile, last updated, eligibility

Source: Arena
Raw value: 1,500
Percentile: 91.1%
Last updated: recent
Eligibility: headline eligible

Parsed from Arena leaderboard dataset row `gpt-5.4-high`. Category: french. Source rank: #24. Votes: 2396. Organization: openai. License: Proprietary.

91.1% percentile inside its fair comparison set

1,500Raw benchmark valueCI 1,486 - 1,514

Text Arena · German

AR · Multilingual · Human

Observed user preference in Arena's Text Arena german leaderboard.

Rank #31 · Source label: gpt-5.4-high

verified runtimeexact alias

Raw row drilldownsource row, percentile, last updated, eligibility

Source: Arena
Raw value: 1,476
Percentile: 89%
Last updated: recent
Eligibility: headline eligible

Parsed from Arena leaderboard dataset row `gpt-5.4-high`. Category: german. Source rank: #33. Votes: 1088. Organization: openai. License: Proprietary.

89% percentile inside its fair comparison set

1,476Raw benchmark valueCI 1,457 - 1,495

Text Arena · Japanese

AR · Multilingual · Human

Observed user preference in Arena's Text Arena japanese leaderboard.

Rank #19 · Source label: gpt-5.4-high

verified runtimeexact alias

Raw row drilldownsource row, percentile, last updated, eligibility

Source: Arena
Raw value: 1,479
Percentile: 92.6%
Last updated: recent
Eligibility: headline eligible

Parsed from Arena leaderboard dataset row `gpt-5.4-high`. Category: japanese. Source rank: #19. Votes: 671. Organization: openai. License: Proprietary.

92.6% percentile inside its fair comparison set

1,479Raw benchmark valueCI 1,454 - 1,504

Text Arena · Korean

AR · Multilingual · Human

Observed user preference in Arena's Text Arena korean leaderboard.

Rank #34 · Source label: gpt-5.4-high

verified runtimeexact alias

Raw row drilldownsource row, percentile, last updated, eligibility

Source: Arena
Raw value: 1,435
Percentile: 86.5%
Last updated: recent
Eligibility: headline eligible

Parsed from Arena leaderboard dataset row `gpt-5.4-high`. Category: korean. Source rank: #36. Votes: 1106. Organization: openai. License: Proprietary.

86.5% percentile inside its fair comparison set

1,435Raw benchmark valueCI 1,415 - 1,454

Text Arena · Non English

AR · Multilingual · Human

Observed user preference in Arena's Text Arena non_english leaderboard.

Rank #29 · Source label: gpt-5.4-high

verified runtimeexact alias

Raw row drilldownsource row, percentile, last updated, eligibility

Source: Arena
Raw value: 1,468
Percentile: 92.6%
Last updated: recent
Eligibility: headline eligible

Parsed from Arena leaderboard dataset row `gpt-5.4-high`. Category: non_english. Source rank: #29. Votes: 35418. Organization: openai. License: Proprietary.

92.6% percentile inside its fair comparison set

1,468Raw benchmark valueCI 1,464 - 1,473

Text Arena · Polish

AR · Multilingual · Human

Observed user preference in Arena's Text Arena polish leaderboard.

Rank #25 · Source label: gpt-5.4-high

verified runtimeexact alias

Raw row drilldownsource row, percentile, last updated, eligibility

Source: Arena
Raw value: 1,482
Percentile: 87.9%
Last updated: recent
Eligibility: headline eligible

Parsed from Arena leaderboard dataset row `gpt-5.4-high`. Category: polish. Source rank: #26. Votes: 1320. Organization: openai. License: Proprietary.

87.9% percentile inside its fair comparison set

1,482Raw benchmark valueCI 1,465 - 1,498

Text Arena · Russian

AR · Multilingual · Human

Observed user preference in Arena's Text Arena russian leaderboard.

Rank #23 · Source label: gpt-5.4-high

verified runtimeexact alias

Raw row drilldownsource row, percentile, last updated, eligibility

Source: Arena
Raw value: 1,487
Percentile: 93.5%
Last updated: recent
Eligibility: headline eligible

Parsed from Arena leaderboard dataset row `gpt-5.4-high`. Category: russian. Source rank: #24. Votes: 7209. Organization: openai. License: Proprietary.

93.5% percentile inside its fair comparison set

1,487Raw benchmark valueCI 1,479 - 1,495

Text Arena · Spanish

AR · Multilingual · Human

Observed user preference in Arena's Text Arena spanish leaderboard.

Rank #45 · Source label: gpt-5.4

verified runtimeexact alias

Raw row drilldownsource row, percentile, last updated, eligibility

Source: Arena
Raw value: 1,461
Percentile: 83%
Last updated: recent
Eligibility: headline eligible

Parsed from Arena leaderboard dataset row `gpt-5.4`. Category: spanish. Source rank: #49. Votes: 2037. Organization: openai. License: Proprietary.

83% percentile inside its fair comparison set

1,461Raw benchmark valueCI 1,446 - 1,475

Text Arena · Chinese · No Style Control

AR · Multilingual · Human

Observed user preference in Arena's Text Arena chinese leaderboard.

Rank #30 · Source label: gpt-5.4-high

verified runtimeexact alias

Raw row drilldownsource row, percentile, last updated, eligibility

Source: Arena
Raw value: 1,520
Percentile: 91.8%
Last updated: recent
Eligibility: headline eligible

Parsed from Arena leaderboard dataset row `gpt-5.4-high`. Category: chinese. Source rank: #32. Votes: 4132. Organization: openai. License: Proprietary.

91.8% percentile inside its fair comparison set

1,520Raw benchmark valueCI 1,510 - 1,530

Text Arena · French · No Style Control

AR · Multilingual · Human

Observed user preference in Arena's Text Arena french leaderboard.

Rank #18 · Source label: gpt-5.4-high

verified runtimeexact alias

Raw row drilldownsource row, percentile, last updated, eligibility

Source: Arena
Raw value: 1,493
Percentile: 93.4%
Last updated: recent
Eligibility: headline eligible

Parsed from Arena leaderboard dataset row `gpt-5.4-high`. Category: french. Source rank: #18. Votes: 2396. Organization: openai. License: Proprietary.

93.4% percentile inside its fair comparison set

1,493Raw benchmark valueCI 1,478 - 1,507

Text Arena · German · No Style Control

AR · Multilingual · Human

Observed user preference in Arena's Text Arena german leaderboard.

Rank #24 · Source label: gpt-5.4-high

verified runtimeexact alias

Raw row drilldownsource row, percentile, last updated, eligibility

Source: Arena
Raw value: 1,472
Percentile: 91.5%
Last updated: recent
Eligibility: headline eligible

Parsed from Arena leaderboard dataset row `gpt-5.4-high`. Category: german. Source rank: #25. Votes: 1088. Organization: openai. License: Proprietary.

91.5% percentile inside its fair comparison set

1,472Raw benchmark valueCI 1,453 - 1,491

Text Arena · Japanese · No Style Control

AR · Multilingual · Human

Observed user preference in Arena's Text Arena japanese leaderboard.

Rank #16 · Source label: gpt-5.4-high

verified runtimeexact alias

Raw row drilldownsource row, percentile, last updated, eligibility

Source: Arena
Raw value: 1,483
Percentile: 93.8%
Last updated: recent
Eligibility: headline eligible

Parsed from Arena leaderboard dataset row `gpt-5.4-high`. Category: japanese. Source rank: #16. Votes: 671. Organization: openai. License: Proprietary.

93.8% percentile inside its fair comparison set

1,483Raw benchmark valueCI 1,459 - 1,508

Text Arena · Korean · No Style Control

AR · Multilingual · Human

Observed user preference in Arena's Text Arena korean leaderboard.

Rank #19 · Source label: gpt-5.4-high

verified runtimeexact alias

Raw row drilldownsource row, percentile, last updated, eligibility

Source: Arena
Raw value: 1,450
Percentile: 92.6%
Last updated: recent
Eligibility: headline eligible

Parsed from Arena leaderboard dataset row `gpt-5.4-high`. Category: korean. Source rank: #19. Votes: 1106. Organization: openai. License: Proprietary.

92.6% percentile inside its fair comparison set

1,450Raw benchmark valueCI 1,430 - 1,469

Text Arena · Non English · No Style Control

AR · Multilingual · Human

Observed user preference in Arena's Text Arena non_english leaderboard.

Rank #25 · Source label: gpt-5.4-high

verified runtimeexact alias

Raw row drilldownsource row, percentile, last updated, eligibility

Source: Arena
Raw value: 1,465
Percentile: 93.6%
Last updated: recent
Eligibility: headline eligible

Parsed from Arena leaderboard dataset row `gpt-5.4-high`. Category: non_english. Source rank: #26. Votes: 35418. Organization: openai. License: Proprietary.

93.6% percentile inside its fair comparison set

1,465Raw benchmark valueCI 1,460 - 1,470

Text Arena · Polish · No Style Control

AR · Multilingual · Human

Observed user preference in Arena's Text Arena polish leaderboard.

Rank #25 · Source label: gpt-5.4-high

verified runtimeexact alias

Raw row drilldownsource row, percentile, last updated, eligibility

Source: Arena
Raw value: 1,471
Percentile: 87.9%
Last updated: recent
Eligibility: headline eligible

Parsed from Arena leaderboard dataset row `gpt-5.4-high`. Category: polish. Source rank: #26. Votes: 1320. Organization: openai. License: Proprietary.

87.9% percentile inside its fair comparison set

1,471Raw benchmark valueCI 1,455 - 1,488

Text Arena · Russian · No Style Control

AR · Multilingual · Human

Observed user preference in Arena's Text Arena russian leaderboard.

Rank #22 · Source label: gpt-5.4-high

verified runtimeexact alias

Raw row drilldownsource row, percentile, last updated, eligibility

Source: Arena
Raw value: 1,480
Percentile: 93.8%
Last updated: recent
Eligibility: headline eligible

Parsed from Arena leaderboard dataset row `gpt-5.4-high`. Category: russian. Source rank: #23. Votes: 7209. Organization: openai. License: Proprietary.

93.8% percentile inside its fair comparison set

1,480Raw benchmark valueCI 1,473 - 1,488

Text Arena · Spanish · No Style Control

AR · Multilingual · Human

Observed user preference in Arena's Text Arena spanish leaderboard.

Rank #43 · Source label: gpt-5.4-high

verified runtimeexact alias

Raw row drilldownsource row, percentile, last updated, eligibility

Source: Arena
Raw value: 1,454
Percentile: 83.8%
Last updated: recent
Eligibility: headline eligible

Parsed from Arena leaderboard dataset row `gpt-5.4-high`. Category: spanish. Source rank: #45. Votes: 2062. Organization: openai. License: Proprietary.

83.8% percentile inside its fair comparison set

1,454Raw benchmark valueCI 1,440 - 1,468

Vision Arena · Chinese

AR · Multilingual · Human

Observed user preference in Arena's Vision Arena chinese leaderboard.

Rank #25 · Source label: gpt-5.4-high

verified runtimeexact alias

Raw row drilldownsource row, percentile, last updated, eligibility

Source: Arena
Raw value: 1,316
Percentile: 76.9%
Last updated: recent
Eligibility: headline eligible

Parsed from Arena leaderboard dataset row `gpt-5.4-high`. Category: chinese. Source rank: #27. Votes: 1768. Organization: openai. License: Proprietary.

76.9% percentile inside its fair comparison set

1,316Raw benchmark valueCI 1,299 - 1,333

Vision Arena · Chinese · No Style Control

AR · Multilingual · Human

Observed user preference in Arena's Vision Arena chinese leaderboard.

Rank #17 · Source label: gpt-5.4-high

verified runtimeexact alias

Raw row drilldownsource row, percentile, last updated, eligibility

Source: Arena
Raw value: 1,353
Percentile: 84.6%
Last updated: recent
Eligibility: headline eligible

Parsed from Arena leaderboard dataset row `gpt-5.4-high`. Category: chinese. Source rank: #19. Votes: 1768. Organization: openai. License: Proprietary.

84.6% percentile inside its fair comparison set

1,353Raw benchmark valueCI 1,336 - 1,370

Source links and registry checks

official

Arena

Oct 3, 2026

source →

official

Arena

Oct 3, 2026

source →

official

Arena

Oct 3, 2026

source →

official

Arena

Oct 3, 2026

source →

official

Arena

Oct 3, 2026

source →

official

Arena

Oct 3, 2026

source →

official

Arena

Oct 3, 2026

source →

official

Arena

Oct 3, 2026

source →

official

Arena

Oct 3, 2026

source →

official

Arena

Oct 3, 2026

source →

official

Arena

Oct 3, 2026

source →

official

Arena

Oct 3, 2026

source →

official

Arena

Oct 3, 2026

source →

official

Arena

Oct 3, 2026

source →

official

Arena

Oct 3, 2026

source →

official

Arena

Oct 3, 2026

source →

official

Arena

Oct 3, 2026

source →

official

Arena

Oct 3, 2026

source →

official

Arena

Oct 3, 2026

source →

official

Arena

Oct 3, 2026

source →

official

Arena

Oct 3, 2026

source →

official

Arena

Oct 3, 2026

source →

official

Arena

Oct 3, 2026

source →

official

Arena

Oct 3, 2026

source →

official

Arena

Oct 3, 2026

source →

official

Arena

Oct 3, 2026

source →

official

Arena

Oct 3, 2026

source →

official

Arena

Oct 3, 2026

source →

official

Arena

Oct 3, 2026

source →

official

Arena

Oct 3, 2026

source →

official

Arena

Oct 3, 2026

source →

official

Arena

Oct 3, 2026

source →

official

Arena

Oct 3, 2026

source →

official

Arena

Oct 3, 2026

source →

official

Arena

Oct 3, 2026

source →

official

Arena

Oct 3, 2026

source →

official

Arena

Oct 3, 2026

source →

official

Arena

Oct 3, 2026

source →

official

Arena

Oct 3, 2026

source →

official

Arena

Oct 3, 2026

source →

official

Arena

Oct 3, 2026

source →

official

Arena

Oct 3, 2026

source →

official

Arena

Oct 3, 2026

source →

official

Arena

Oct 3, 2026

source →

official

Artificial Analysis

Oct 3, 2026

source →

official

LiveBench

Oct 3, 2026

source →

official

Scale Labs

Oct 3, 2026

source →

official

Scale Labs

Oct 3, 2026

source →

official

Scale Labs

Oct 3, 2026

source →

official

Scale Labs

Oct 3, 2026

source →

official

Scale Labs

Oct 3, 2026

source →

official

Scale Labs

Oct 3, 2026

source →

official

Scale Labs

Oct 3, 2026

source →

GPT-5.4

Current snapshot.

Source-linked scores by benchmark

Source links and registry checks

Loading model evidence.

GPT-5.4

Current snapshot.

Source-linked scores by benchmark

Source links and registry checks