source
Arena
Benchmarks · /benchmarks/arena-document
Document Arena
Blind preference arena for document-heavy prompts, PDFs, and long-form file understanding.
Passport
Visible tradeoffsThis is a human preference signal, so it tells you what people liked side by side, not what is formally correct.
metric
Arena rating (rating)
judge
Human
direction
higher better
group id
arena_document_2026_q2
domain
Document understanding
What it measures vs what it misses
✓ Measures
Observed user preference on document-grounded outputs. How often a model wins when users compare file-aware responses head to head.
✗ Misses
Objective document extraction accuracy. Fine-grained breakdowns across OCR, tables, charts, and legal or financial workflows.
Why this countsIt matters when the job is reading PDFs, tables, forms, or mixed-layout documents rather than plain chat.Comparable-group ruleThis percentile only compares models inside the exact benchmark/version group shown here. It is not a universal score.What it missesIt does not stand in for end-to-end document workflow quality or search relevance.
Leaderboard · this benchmark version
#1 · Claude Opus 4.6
AR · May 13, 2026
1,513
#2 · Claude Opus 4.7
AR · May 13, 2026
1,509
#3 · Claude Sonnet 4.6
AR · May 13, 2026
1,495
#4 · GPT-5.5
AR · May 13, 2026
1,492
#5 · GPT-5.4
AR · May 13, 2026
1,474
#6 · Claude Opus 4.5
AR · May 13, 2026
1,466
#7 · kimi-k2.6
AR · May 13, 2026
1,454
#8 · muse-spark
AR · May 13, 2026
1,452
#9 · Claude Sonnet 4.5
AR · May 13, 2026
1,450
#10 · Gemini 3.1 Pro Preview
AR · May 13, 2026
1,443
#11 · Gemini 3 Pro Preview
AR · May 13, 2026
1,439
#12 · kimi-k2.5-thinking
AR · May 13, 2026
1,437
#13 · Gemini 2.5 Pro
AR · May 13, 2026
1,427
#14 · gemma-4-31b
AR · May 13, 2026
1,424
#15 · Claude Haiku 4.5
AR · May 13, 2026
1,423
#16 · Grok 4.20
AR · May 13, 2026
1,420
#17 · Gemini 3 Flash
AR · May 13, 2026
1,418
#18 · GPT-5.2
AR · May 13, 2026
1,407
#19 · gpt-5.5-instant
AR · May 13, 2026
1,407
#20 · GPT-5.1
AR · May 13, 2026
1,407