Benchmarks · /benchmarks/terminal-bench-4

Terminal-Bench 4.0

Name: Terminal-Bench 4.0
Creator: Terminal-Bench

Official published Terminal-Bench 4.0 leaderboard for realistic multi-step terminal tasks.

Source · Terminal-Bench
Version · terminal-bench snapshot 2026-10-03
Scores · 27

Test details

Verified but agingThis is an objective signal, so it is mainly about measurable task performance rather than public taste.

source

Terminal-Bench

metric

Accuracy (%)

judge

Objective

direction

higher better

group id

terminal_bench_4_live

domain

Coding

What it measures vs what it misses

✓ Measures

End-to-end task success on the Terminal-Bench 4.0 task set with the published agent scaffold and reasoning configuration.

✗ Misses

Scores from earlier Terminal-Bench versions use different tasks and cannot be compared directly. IDE-native workflows, code review quality, and non-terminal product engineering work.

Why this countsIt tells you whether the model can generate, repair, and reason over code under evaluator pressure rather than marketing examples.Same-test ruleThis percentile only compares models inside the exact benchmark/version group shown here. It is not a universal score.What it missesIt does not fully capture repo-scale iteration, IDE ergonomics, or long debugging loops.

Leaderboard · this benchmark version

#1 · gpt-6-astra-max

TERMINAL-BENCH · Sep 3, 2026

Source label: GPT-6 Astra (max)

verified runtimeexact alias

Raw row drilldownsource, percentile, eligibility

Source URL: https://www.tbench.ai/
Percentile: 100%
Last updated: recent
Eligibility: headline eligible
Identity: provider alias (0.92)

Parsed from the public Terminal-Bench 4.0 official published leaderboard (status: display). Collapse policy: highest verified score per canonical model. Selected agent: Codex. Display model: GPT-6 Astra. Agent URL: https://openai.com/codex/. Reported stderr: 1.423 percentage points.

58.2%55.4% - 61%

#2 · claude-fable-5.1-max

TERMINAL-BENCH · Sep 1, 2026

Source label: Fable 5.1 (max)

verified runtimeexact direct

Raw row drilldownsource, percentile, eligibility

Source URL: https://www.tbench.ai/
Percentile: 96.2%
Last updated: recent
Eligibility: headline eligible
Identity: exact (1.00)

Parsed from the public Terminal-Bench 4.0 official published leaderboard (status: display). Collapse policy: highest verified score per canonical model. Selected agent: Claude Code. Display model: Fable 5.1. Agent URL: https://claude.com/product/claude-code. Reported stderr: 1.918 percentage points.

57.9%54.1% - 61.6%

#3 · Claude Fable 5.1 (Adaptive Reasoning, Xhigh Effort, Default Fallback)

TERMINAL-BENCH · Sep 1, 2026

Source label: Fable 5.1 (xhigh)

verified runtimeexact direct

Raw row drilldownsource, percentile, eligibility

Source URL: https://www.tbench.ai/
Percentile: 96.2%
Last updated: recent
Eligibility: headline eligible
Identity: exact (1.00)

57.9%54.5% - 61.2%

#4 · GPT-6 Astra (High)

TERMINAL-BENCH · Sep 3, 2026

Source label: GPT-6 Astra (high)

verified runtimeexact direct

Raw row drilldownsource, percentile, eligibility

Source URL: https://www.tbench.ai/
Percentile: 96.2%
Last updated: recent
Eligibility: headline eligible
Identity: exact (1.00)

57.9%54.9% - 60.8%

#5 · GPT-6 Astra (Xhigh)

TERMINAL-BENCH · Sep 3, 2026

Source label: GPT-6 Astra (xhigh)

verified runtimeexact direct

Raw row drilldownsource, percentile, eligibility

Source URL: https://www.tbench.ai/
Percentile: 96.2%
Last updated: recent
Eligibility: headline eligible
Identity: exact (1.00)

57.9%55.2% - 60.6%

#6 · Claude Fable 5.1 (Adaptive Reasoning, High Effort, Default Fallback)

TERMINAL-BENCH · Sep 1, 2026

Source label: Fable 5.1 (high)

verified runtimeexact direct

Raw row drilldownsource, percentile, eligibility

Source URL: https://www.tbench.ai/
Percentile: 80.8%
Last updated: recent
Eligibility: headline eligible
Identity: exact (1.00)

54.6%51.1% - 58%

#7 · GPT-6 Astra (Medium)

TERMINAL-BENCH · Sep 3, 2026

Source label: GPT-6 Astra (medium)

verified runtimeexact direct

Raw row drilldownsource, percentile, eligibility

Source URL: https://www.tbench.ai/
Percentile: 76.9%
Last updated: recent
Eligibility: headline eligible
Identity: exact (1.00)

54.2%51.6% - 56.9%

#8 · Claude Fable 5.1 (Adaptive Reasoning, Medium Effort, Default Fallback)

TERMINAL-BENCH · Sep 1, 2026

Source label: Fable 5.1 (medium)

verified runtimeexact direct

Raw row drilldownsource, percentile, eligibility

Source URL: https://www.tbench.ai/
Percentile: 73.1%
Last updated: recent
Eligibility: headline eligible
Identity: exact (1.00)

53.9%50.6% - 57.3%

#9 · Claude Opus 5 (Adaptive Reasoning, Xhigh Effort)

TERMINAL-BENCH · Jul 24, 2026

Source label: Opus 5 (xhigh)

verified runtimeexact direct

Raw row drilldownsource, percentile, eligibility

Source URL: https://www.tbench.ai/
Percentile: 73.1%
Last updated: aging
Eligibility: headline eligible
Identity: exact (1.00)

Parsed from the public Terminal-Bench 4.0 official published leaderboard (status: display). Collapse policy: highest verified score per canonical model. Selected agent: Claude Code. Display model: Opus 5. Agent URL: https://claude.com/product/claude-code. Reported stderr: 1.617 percentage points.

53.9%50.8% - 57.1%

#10 · claude-opus-5-max

TERMINAL-BENCH · Jul 24, 2026

Source label: Opus 5 (max)

verified runtimeexact direct

Raw row drilldownsource, percentile, eligibility

Source URL: https://www.tbench.ai/
Percentile: 65.4%
Last updated: aging
Eligibility: headline eligible
Identity: exact (1.00)

Parsed from the public Terminal-Bench 4.0 official published leaderboard (status: display). Collapse policy: highest verified score per canonical model. Selected agent: Claude Code. Display model: Opus 5. Agent URL: https://claude.com/product/claude-code. Reported stderr: 1.730 percentage points.

51.8%48.4% - 55.2%

#11 · GPT-6 Astra (Low)

TERMINAL-BENCH · Sep 3, 2026

Source label: GPT-6 Astra (low)

verified runtimeexact direct

Raw row drilldownsource, percentile, eligibility

Source URL: https://www.tbench.ai/
Percentile: 61.5%
Last updated: recent
Eligibility: headline eligible
Identity: exact (1.00)

50.6%47.9% - 53.4%

#12 · claude-opus-5-high

TERMINAL-BENCH · Jul 24, 2026

Source label: Opus 5 (high)

verified runtimeexact direct

Raw row drilldownsource, percentile, eligibility

Source URL: https://www.tbench.ai/
Percentile: 57.7%
Last updated: aging
Eligibility: headline eligible
Identity: exact (1.00)

Parsed from the public Terminal-Bench 4.0 official published leaderboard (status: display). Collapse policy: highest verified score per canonical model. Selected agent: Claude Code. Display model: Opus 5. Agent URL: https://claude.com/product/claude-code. Reported stderr: 1.903 percentage points.

50.3%46.6% - 54%

#13 · Claude Opus 5 (Adaptive Reasoning, Medium Effort)

TERMINAL-BENCH · Jul 24, 2026

Source label: Opus 5 (medium)

verified runtimeexact direct

Raw row drilldownsource, percentile, eligibility

Source URL: https://www.tbench.ai/
Percentile: 53.8%
Last updated: aging
Eligibility: headline eligible
Identity: exact (1.00)

Parsed from the public Terminal-Bench 4.0 official published leaderboard (status: display). Collapse policy: highest verified score per canonical model. Selected agent: Claude Code. Display model: Opus 5. Agent URL: https://claude.com/product/claude-code. Reported stderr: 1.954 percentage points.

44.9%41% - 48.7%

#14 · Claude Fable 5

TERMINAL-BENCH · Jun 9, 2026

Source label: Fable 5 (max)

verified runtimeexact alias

Raw row drilldownsource, percentile, eligibility

Source URL: https://www.tbench.ai/
Percentile: 50%
Last updated: stale
Eligibility: headline eligible
Identity: provider alias (0.92)

Parsed from the public Terminal-Bench 4.0 official published leaderboard (status: display). Collapse policy: highest verified score per canonical model. Selected agent: Claude Code. Display model: Fable 5. Agent URL: https://claude.com/product/claude-code. Reported stderr: 1.964 percentage points.

44.6%40.7% - 48.4%

#15 · Claude Fable 5.1 (Adaptive Reasoning, Low Effort, Default Fallback)

TERMINAL-BENCH · Sep 1, 2026

Source label: Fable 5.1 (low)

verified runtimeexact direct

Raw row drilldownsource, percentile, eligibility

Source URL: https://www.tbench.ai/
Percentile: 46.2%
Last updated: recent
Eligibility: headline eligible
Identity: exact (1.00)

43.3%39.7% - 46.9%

#16 · glm-5.3-max

TERMINAL-BENCH · Aug 14, 2026

Source label: GLM-5.3 (max)

verified runtimeexact alias

Raw row drilldownsource, percentile, eligibility

Source URL: https://www.tbench.ai/
Percentile: 42.3%
Last updated: aging
Eligibility: headline eligible
Identity: provider alias (0.92)

Parsed from the public Terminal-Bench 4.0 official published leaderboard (status: display). Collapse policy: highest verified score per canonical model. Selected agent: Claude Code. Display model: GLM-5.3. Agent URL: https://claude.com/product/claude-code. Reported stderr: 1.648 percentage points.

41.8%38.6% - 45.1%

#17 · grok-4.7-xhigh

TERMINAL-BENCH · Sep 21, 2026

Source label: Grok 4.7 (xhigh)

verified runtimeexact alias

Raw row drilldownsource, percentile, eligibility

Source URL: https://www.tbench.ai/
Percentile: 38.5%
Last updated: recent
Eligibility: headline eligible
Identity: provider alias (0.92)

Parsed from the public Terminal-Bench 4.0 official published leaderboard (status: display). Collapse policy: highest verified score per canonical model. Selected agent: Grok Build. Display model: Grok 4.7. Agent URL: https://x.ai/build. Reported stderr: 1.806 percentage points.

37.6%34% - 41.1%

#18 · GPT-5.6 Sol

TERMINAL-BENCH · Jun 26, 2026

Source label: GPT-5.6 Sol (max)

verified runtimeexact alias

Raw row drilldownsource, percentile, eligibility

Source URL: https://www.tbench.ai/
Percentile: 34.6%
Last updated: stale
Eligibility: headline eligible
Identity: provider alias (0.92)

Parsed from the public Terminal-Bench 4.0 official published leaderboard (status: display). Collapse policy: highest verified score per canonical model. Selected agent: Codex. Display model: GPT-5.6 Sol. Agent URL: https://openai.com/codex/. Reported stderr: 1.929 percentage points.

37.3%33.5% - 41.1%

#19 · Claude Opus 5 (Adaptive Reasoning, Low Effort)

TERMINAL-BENCH · Jul 24, 2026

Source label: Opus 5 (low)

verified runtimeexact direct

Raw row drilldownsource, percentile, eligibility

Source URL: https://www.tbench.ai/
Percentile: 30.8%
Last updated: aging
Eligibility: headline eligible
Identity: exact (1.00)

Parsed from the public Terminal-Bench 4.0 official published leaderboard (status: display). Collapse policy: highest verified score per canonical model. Selected agent: Claude Code. Display model: Opus 5. Agent URL: https://claude.com/product/claude-code. Reported stderr: 2.010 percentage points.

34.9%30.9% - 38.8%

#20 · Claude Opus 4.8

TERMINAL-BENCH · May 28, 2026

Source label: Opus 4.8 (max)

verified runtimeexact alias

Raw row drilldownsource, percentile, eligibility

Source URL: https://www.tbench.ai/
Percentile: 26.9%
Last updated: archived
Eligibility: headline eligible
Identity: provider alias (0.92)

Parsed from the public Terminal-Bench 4.0 official published leaderboard (status: display). Collapse policy: highest verified score per canonical model. Selected agent: Claude Code. Display model: Opus 4.8. Agent URL: https://claude.com/product/claude-code. Reported stderr: 1.816 percentage points.

23.6%20.1% - 27.2%

#21 · GPT-5.6 Terra

TERMINAL-BENCH · Jun 26, 2026

Source label: GPT-5.6 Terra (max)

verified runtimeexact alias

Raw row drilldownsource, percentile, eligibility

Source URL: https://www.tbench.ai/
Percentile: 23.1%
Last updated: stale
Eligibility: headline eligible
Identity: provider alias (0.92)

Parsed from the public Terminal-Bench 4.0 official published leaderboard (status: display). Collapse policy: highest verified score per canonical model. Selected agent: Codex. Display model: GPT-5.6 Terra. Agent URL: https://openai.com/codex/. Reported stderr: 1.658 percentage points.

21.5%18.3% - 24.8%

#22 · grok-4.6-high

TERMINAL-BENCH · Aug 12, 2026

Source label: Grok 4.6 (high)

verified runtimeexact alias

Raw row drilldownsource, percentile, eligibility

Source URL: https://www.tbench.ai/
Percentile: 19.2%
Last updated: aging
Eligibility: headline eligible
Identity: provider alias (0.92)

Parsed from the public Terminal-Bench 4.0 official published leaderboard (status: display). Collapse policy: highest verified score per canonical model. Selected agent: Grok Build. Display model: Grok 4.6. Agent URL: https://x.ai/build. Reported stderr: 1.577 percentage points.

20.3%17.2% - 23.4%

#23 · gemini-3.8-flash-high

TERMINAL-BENCH · Sep 2, 2026

Source label: Gemini 3.8 Flash (high)

verified runtimeexact alias

Raw row drilldownsource, percentile, eligibility

Source URL: https://www.tbench.ai/
Percentile: 15.4%
Last updated: recent
Eligibility: headline eligible
Identity: provider alias (0.92)

Parsed from the public Terminal-Bench 4.0 official published leaderboard (status: display). Collapse policy: highest verified score per canonical model. Selected agent: mini-SWE-agent. Display model: Gemini 3.8 Flash. Agent URL: https://github.com/SWE-agent/mini-swe-agent. Reported stderr: 1.714 percentage points.

19.1%15.7% - 22.5%

#24 · GPT-5.6 Luna

TERMINAL-BENCH · Jun 26, 2026

Source label: GPT-5.6 Luna (max)

verified runtimeexact alias

Raw row drilldownsource, percentile, eligibility

Source URL: https://www.tbench.ai/
Percentile: 11.5%
Last updated: stale
Eligibility: headline eligible
Identity: provider alias (0.92)

Parsed from the public Terminal-Bench 4.0 official published leaderboard (status: display). Collapse policy: highest verified score per canonical model. Selected agent: Codex. Display model: GPT-5.6 Luna. Agent URL: https://openai.com/codex/. Reported stderr: 1.454 percentage points.

17.3%14.4% - 20.1%

#25 · Claude Sonnet 5 (Adaptive Reasoning, Max Effort)

TERMINAL-BENCH · Jun 30, 2026

Source label: Sonnet 5 (max)

verified runtimeexact alias

Raw row drilldownsource, percentile, eligibility

Source URL: https://www.tbench.ai/
Percentile: 7.7%
Last updated: stale
Eligibility: headline eligible
Identity: provider alias (0.92)

Parsed from the public Terminal-Bench 4.0 official published leaderboard (status: display). Collapse policy: highest verified score per canonical model. Selected agent: Claude Code. Display model: Sonnet 5. Agent URL: https://claude.com/product/claude-code. Reported stderr: 1.561 percentage points.

12.4%9.4% - 15.5%

#26 · grok-4.5

TERMINAL-BENCH · Jul 16, 2026

Source label: Grok 4.5 (high)

verified runtimeexact alias

Raw row drilldownsource, percentile, eligibility

Source URL: https://www.tbench.ai/
Percentile: 7.7%
Last updated: aging
Eligibility: headline eligible
Identity: provider alias (0.92)

Parsed from the public Terminal-Bench 4.0 official published leaderboard (status: display). Collapse policy: highest verified score per canonical model. Selected agent: Grok Build. Display model: Grok 4.5. Agent URL: https://x.ai/build. Reported stderr: 1.337 percentage points.

12.4%9.8% - 15%

#27 · gemini-3.7-flash-high

TERMINAL-BENCH · Aug 13, 2026

Source label: Gemini 3.7 Flash (high)

verified runtimeexact alias

Raw row drilldownsource, percentile, eligibility

Source URL: https://www.tbench.ai/
Percentile: 0%
Last updated: aging
Eligibility: headline eligible
Identity: provider alias (0.92)

Parsed from the public Terminal-Bench 4.0 official published leaderboard (status: display). Collapse policy: highest verified score per canonical model. Selected agent: mini-SWE-agent. Display model: Gemini 3.7 Flash. Agent URL: https://github.com/SWE-agent/mini-swe-agent. Reported stderr: 1.250 percentage points.

11.2%8.8% - 13.7%

Benchmarks · /benchmarks/terminal-bench-4

Terminal-Bench 4.0

Official published Terminal-Bench 4.0 leaderboard for realistic multi-step terminal tasks.

Source · Terminal-Bench
Version · terminal-bench snapshot 2026-10-03
Scores · 27

Test details

Verified but agingThis is an objective signal, so it is mainly about measurable task performance rather than public taste.

source

Terminal-Bench

metric

Accuracy (%)

judge

Objective

direction

higher better

group id

terminal_bench_4_live

domain

Coding

What it measures vs what it misses

✓ Measures

End-to-end task success on the Terminal-Bench 4.0 task set with the published agent scaffold and reasoning configuration.

✗ Misses

Scores from earlier Terminal-Bench versions use different tasks and cannot be compared directly. IDE-native workflows, code review quality, and non-terminal product engineering work.

Leaderboard · this benchmark version

#1 · gpt-6-astra-max

TERMINAL-BENCH · Sep 3, 2026

Source label: GPT-6 Astra (max)

verified runtimeexact alias

Raw row drilldownsource, percentile, eligibility

Source URL: https://www.tbench.ai/
Percentile: 100%
Last updated: recent
Eligibility: headline eligible
Identity: provider alias (0.92)

58.2%55.4% - 61%

#2 · claude-fable-5.1-max

TERMINAL-BENCH · Sep 1, 2026

Source label: Fable 5.1 (max)

verified runtimeexact direct

Raw row drilldownsource, percentile, eligibility

Source URL: https://www.tbench.ai/
Percentile: 96.2%
Last updated: recent
Eligibility: headline eligible
Identity: exact (1.00)

57.9%54.1% - 61.6%

#3 · Claude Fable 5.1 (Adaptive Reasoning, Xhigh Effort, Default Fallback)

TERMINAL-BENCH · Sep 1, 2026

Source label: Fable 5.1 (xhigh)

verified runtimeexact direct

Raw row drilldownsource, percentile, eligibility

Source URL: https://www.tbench.ai/
Percentile: 96.2%
Last updated: recent
Eligibility: headline eligible
Identity: exact (1.00)

57.9%54.5% - 61.2%

#4 · GPT-6 Astra (High)

TERMINAL-BENCH · Sep 3, 2026

Source label: GPT-6 Astra (high)

verified runtimeexact direct

Raw row drilldownsource, percentile, eligibility

Source URL: https://www.tbench.ai/
Percentile: 96.2%
Last updated: recent
Eligibility: headline eligible
Identity: exact (1.00)

57.9%54.9% - 60.8%

#5 · GPT-6 Astra (Xhigh)

TERMINAL-BENCH · Sep 3, 2026

Source label: GPT-6 Astra (xhigh)

verified runtimeexact direct

Raw row drilldownsource, percentile, eligibility

Source URL: https://www.tbench.ai/
Percentile: 96.2%
Last updated: recent
Eligibility: headline eligible
Identity: exact (1.00)

57.9%55.2% - 60.6%

#6 · Claude Fable 5.1 (Adaptive Reasoning, High Effort, Default Fallback)

TERMINAL-BENCH · Sep 1, 2026

Source label: Fable 5.1 (high)

verified runtimeexact direct

Raw row drilldownsource, percentile, eligibility

Source URL: https://www.tbench.ai/
Percentile: 80.8%
Last updated: recent
Eligibility: headline eligible
Identity: exact (1.00)

54.6%51.1% - 58%

#7 · GPT-6 Astra (Medium)

TERMINAL-BENCH · Sep 3, 2026

Source label: GPT-6 Astra (medium)

verified runtimeexact direct

Raw row drilldownsource, percentile, eligibility

Source URL: https://www.tbench.ai/
Percentile: 76.9%
Last updated: recent
Eligibility: headline eligible
Identity: exact (1.00)

54.2%51.6% - 56.9%

#8 · Claude Fable 5.1 (Adaptive Reasoning, Medium Effort, Default Fallback)

TERMINAL-BENCH · Sep 1, 2026

Source label: Fable 5.1 (medium)

verified runtimeexact direct

Raw row drilldownsource, percentile, eligibility

Source URL: https://www.tbench.ai/
Percentile: 73.1%
Last updated: recent
Eligibility: headline eligible
Identity: exact (1.00)

53.9%50.6% - 57.3%

#9 · Claude Opus 5 (Adaptive Reasoning, Xhigh Effort)

TERMINAL-BENCH · Jul 24, 2026

Source label: Opus 5 (xhigh)

verified runtimeexact direct

Raw row drilldownsource, percentile, eligibility

Source URL: https://www.tbench.ai/
Percentile: 73.1%
Last updated: aging
Eligibility: headline eligible
Identity: exact (1.00)

Parsed from the public Terminal-Bench 4.0 official published leaderboard (status: display). Collapse policy: highest verified score per canonical model. Selected agent: Claude Code. Display model: Opus 5. Agent URL: https://claude.com/product/claude-code. Reported stderr: 1.617 percentage points.

53.9%50.8% - 57.1%

#10 · claude-opus-5-max

TERMINAL-BENCH · Jul 24, 2026

Source label: Opus 5 (max)

verified runtimeexact direct

Raw row drilldownsource, percentile, eligibility

Source URL: https://www.tbench.ai/
Percentile: 65.4%
Last updated: aging
Eligibility: headline eligible
Identity: exact (1.00)

Parsed from the public Terminal-Bench 4.0 official published leaderboard (status: display). Collapse policy: highest verified score per canonical model. Selected agent: Claude Code. Display model: Opus 5. Agent URL: https://claude.com/product/claude-code. Reported stderr: 1.730 percentage points.

51.8%48.4% - 55.2%

#11 · GPT-6 Astra (Low)

TERMINAL-BENCH · Sep 3, 2026

Source label: GPT-6 Astra (low)

verified runtimeexact direct

Raw row drilldownsource, percentile, eligibility

Source URL: https://www.tbench.ai/
Percentile: 61.5%
Last updated: recent
Eligibility: headline eligible
Identity: exact (1.00)

50.6%47.9% - 53.4%

#12 · claude-opus-5-high

TERMINAL-BENCH · Jul 24, 2026

Source label: Opus 5 (high)

verified runtimeexact direct

Raw row drilldownsource, percentile, eligibility

Source URL: https://www.tbench.ai/
Percentile: 57.7%
Last updated: aging
Eligibility: headline eligible
Identity: exact (1.00)

Parsed from the public Terminal-Bench 4.0 official published leaderboard (status: display). Collapse policy: highest verified score per canonical model. Selected agent: Claude Code. Display model: Opus 5. Agent URL: https://claude.com/product/claude-code. Reported stderr: 1.903 percentage points.

50.3%46.6% - 54%

#13 · Claude Opus 5 (Adaptive Reasoning, Medium Effort)

TERMINAL-BENCH · Jul 24, 2026

Source label: Opus 5 (medium)

verified runtimeexact direct

Raw row drilldownsource, percentile, eligibility

Source URL: https://www.tbench.ai/
Percentile: 53.8%
Last updated: aging
Eligibility: headline eligible
Identity: exact (1.00)

Parsed from the public Terminal-Bench 4.0 official published leaderboard (status: display). Collapse policy: highest verified score per canonical model. Selected agent: Claude Code. Display model: Opus 5. Agent URL: https://claude.com/product/claude-code. Reported stderr: 1.954 percentage points.

44.9%41% - 48.7%

#14 · Claude Fable 5

TERMINAL-BENCH · Jun 9, 2026

Source label: Fable 5 (max)

verified runtimeexact alias

Raw row drilldownsource, percentile, eligibility

Source URL: https://www.tbench.ai/
Percentile: 50%
Last updated: stale
Eligibility: headline eligible
Identity: provider alias (0.92)

Parsed from the public Terminal-Bench 4.0 official published leaderboard (status: display). Collapse policy: highest verified score per canonical model. Selected agent: Claude Code. Display model: Fable 5. Agent URL: https://claude.com/product/claude-code. Reported stderr: 1.964 percentage points.

44.6%40.7% - 48.4%

#15 · Claude Fable 5.1 (Adaptive Reasoning, Low Effort, Default Fallback)

TERMINAL-BENCH · Sep 1, 2026

Source label: Fable 5.1 (low)

verified runtimeexact direct

Raw row drilldownsource, percentile, eligibility

Source URL: https://www.tbench.ai/
Percentile: 46.2%
Last updated: recent
Eligibility: headline eligible
Identity: exact (1.00)

43.3%39.7% - 46.9%

#16 · glm-5.3-max

TERMINAL-BENCH · Aug 14, 2026

Source label: GLM-5.3 (max)

verified runtimeexact alias

Raw row drilldownsource, percentile, eligibility

Source URL: https://www.tbench.ai/
Percentile: 42.3%
Last updated: aging
Eligibility: headline eligible
Identity: provider alias (0.92)

Parsed from the public Terminal-Bench 4.0 official published leaderboard (status: display). Collapse policy: highest verified score per canonical model. Selected agent: Claude Code. Display model: GLM-5.3. Agent URL: https://claude.com/product/claude-code. Reported stderr: 1.648 percentage points.

41.8%38.6% - 45.1%

#17 · grok-4.7-xhigh

TERMINAL-BENCH · Sep 21, 2026

Source label: Grok 4.7 (xhigh)

verified runtimeexact alias

Raw row drilldownsource, percentile, eligibility

Source URL: https://www.tbench.ai/
Percentile: 38.5%
Last updated: recent
Eligibility: headline eligible
Identity: provider alias (0.92)

37.6%34% - 41.1%

#18 · GPT-5.6 Sol

TERMINAL-BENCH · Jun 26, 2026

Source label: GPT-5.6 Sol (max)

verified runtimeexact alias

Raw row drilldownsource, percentile, eligibility

Source URL: https://www.tbench.ai/
Percentile: 34.6%
Last updated: stale
Eligibility: headline eligible
Identity: provider alias (0.92)

Parsed from the public Terminal-Bench 4.0 official published leaderboard (status: display). Collapse policy: highest verified score per canonical model. Selected agent: Codex. Display model: GPT-5.6 Sol. Agent URL: https://openai.com/codex/. Reported stderr: 1.929 percentage points.

37.3%33.5% - 41.1%

#19 · Claude Opus 5 (Adaptive Reasoning, Low Effort)

TERMINAL-BENCH · Jul 24, 2026

Source label: Opus 5 (low)

verified runtimeexact direct

Raw row drilldownsource, percentile, eligibility

Source URL: https://www.tbench.ai/
Percentile: 30.8%
Last updated: aging
Eligibility: headline eligible
Identity: exact (1.00)

Parsed from the public Terminal-Bench 4.0 official published leaderboard (status: display). Collapse policy: highest verified score per canonical model. Selected agent: Claude Code. Display model: Opus 5. Agent URL: https://claude.com/product/claude-code. Reported stderr: 2.010 percentage points.

34.9%30.9% - 38.8%

#20 · Claude Opus 4.8

TERMINAL-BENCH · May 28, 2026

Source label: Opus 4.8 (max)

verified runtimeexact alias

Raw row drilldownsource, percentile, eligibility

Source URL: https://www.tbench.ai/
Percentile: 26.9%
Last updated: archived
Eligibility: headline eligible
Identity: provider alias (0.92)

Parsed from the public Terminal-Bench 4.0 official published leaderboard (status: display). Collapse policy: highest verified score per canonical model. Selected agent: Claude Code. Display model: Opus 4.8. Agent URL: https://claude.com/product/claude-code. Reported stderr: 1.816 percentage points.

23.6%20.1% - 27.2%

#21 · GPT-5.6 Terra

TERMINAL-BENCH · Jun 26, 2026

Source label: GPT-5.6 Terra (max)

verified runtimeexact alias

Raw row drilldownsource, percentile, eligibility

Source URL: https://www.tbench.ai/
Percentile: 23.1%
Last updated: stale
Eligibility: headline eligible
Identity: provider alias (0.92)

Parsed from the public Terminal-Bench 4.0 official published leaderboard (status: display). Collapse policy: highest verified score per canonical model. Selected agent: Codex. Display model: GPT-5.6 Terra. Agent URL: https://openai.com/codex/. Reported stderr: 1.658 percentage points.

21.5%18.3% - 24.8%

#22 · grok-4.6-high

TERMINAL-BENCH · Aug 12, 2026

Source label: Grok 4.6 (high)

verified runtimeexact alias

Raw row drilldownsource, percentile, eligibility

Source URL: https://www.tbench.ai/
Percentile: 19.2%
Last updated: aging
Eligibility: headline eligible
Identity: provider alias (0.92)

Parsed from the public Terminal-Bench 4.0 official published leaderboard (status: display). Collapse policy: highest verified score per canonical model. Selected agent: Grok Build. Display model: Grok 4.6. Agent URL: https://x.ai/build. Reported stderr: 1.577 percentage points.

20.3%17.2% - 23.4%

#23 · gemini-3.8-flash-high

TERMINAL-BENCH · Sep 2, 2026

Source label: Gemini 3.8 Flash (high)

verified runtimeexact alias

Raw row drilldownsource, percentile, eligibility

Source URL: https://www.tbench.ai/
Percentile: 15.4%
Last updated: recent
Eligibility: headline eligible
Identity: provider alias (0.92)

19.1%15.7% - 22.5%

#24 · GPT-5.6 Luna

TERMINAL-BENCH · Jun 26, 2026

Source label: GPT-5.6 Luna (max)

verified runtimeexact alias

Raw row drilldownsource, percentile, eligibility

Source URL: https://www.tbench.ai/
Percentile: 11.5%
Last updated: stale
Eligibility: headline eligible
Identity: provider alias (0.92)

Parsed from the public Terminal-Bench 4.0 official published leaderboard (status: display). Collapse policy: highest verified score per canonical model. Selected agent: Codex. Display model: GPT-5.6 Luna. Agent URL: https://openai.com/codex/. Reported stderr: 1.454 percentage points.

17.3%14.4% - 20.1%

#25 · Claude Sonnet 5 (Adaptive Reasoning, Max Effort)

TERMINAL-BENCH · Jun 30, 2026

Source label: Sonnet 5 (max)

verified runtimeexact alias

Raw row drilldownsource, percentile, eligibility

Source URL: https://www.tbench.ai/
Percentile: 7.7%
Last updated: stale
Eligibility: headline eligible
Identity: provider alias (0.92)

Parsed from the public Terminal-Bench 4.0 official published leaderboard (status: display). Collapse policy: highest verified score per canonical model. Selected agent: Claude Code. Display model: Sonnet 5. Agent URL: https://claude.com/product/claude-code. Reported stderr: 1.561 percentage points.

12.4%9.4% - 15.5%

#26 · grok-4.5

TERMINAL-BENCH · Jul 16, 2026

Source label: Grok 4.5 (high)

verified runtimeexact alias

Raw row drilldownsource, percentile, eligibility

Source URL: https://www.tbench.ai/
Percentile: 7.7%
Last updated: aging
Eligibility: headline eligible
Identity: provider alias (0.92)

Parsed from the public Terminal-Bench 4.0 official published leaderboard (status: display). Collapse policy: highest verified score per canonical model. Selected agent: Grok Build. Display model: Grok 4.5. Agent URL: https://x.ai/build. Reported stderr: 1.337 percentage points.

12.4%9.8% - 15%

#27 · gemini-3.7-flash-high

TERMINAL-BENCH · Aug 13, 2026

Source label: Gemini 3.7 Flash (high)

verified runtimeexact alias

Raw row drilldownsource, percentile, eligibility

Source URL: https://www.tbench.ai/
Percentile: 0%
Last updated: aging
Eligibility: headline eligible
Identity: provider alias (0.92)

Parsed from the public Terminal-Bench 4.0 official published leaderboard (status: display). Collapse policy: highest verified score per canonical model. Selected agent: mini-SWE-agent. Display model: Gemini 3.7 Flash. Agent URL: https://github.com/SWE-agent/mini-swe-agent. Reported stderr: 1.250 percentage points.

11.2%8.8% - 13.7%

Terminal-Bench 4.0

Test details

What it measures vs what it misses

✓ Measures

✗ Misses

Leaderboard · this benchmark version

Loading benchmark evidence.

Terminal-Bench 4.0

Test details

What it measures vs what it misses

✓ Measures

✗ Misses

Leaderboard · this benchmark version