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 1, 2026
1,520
#2 · Claude Opus 4.7
AR · May 1, 2026
1,511
#3 · Claude Sonnet 4.6
AR · May 1, 2026
1,500
#4 · GPT-5.5
AR · May 1, 2026
1,487
#5 · GPT-5.4
AR · May 1, 2026
1,480
#6 · Claude Opus 4.5
AR · May 1, 2026
1,470
#7 · Kimi K2.6
AR · May 1, 2026
1,457
#8 · muse-spark
AR · May 1, 2026
1,457
#9 · Claude Sonnet 4.5
AR · May 1, 2026
1,450
#10 · Gemini 3.1 Pro Preview
AR · May 1, 2026
1,449
#11 · kimi-k2.5-thinking
AR · May 1, 2026
1,444
#12 · Gemini 3 Pro Preview
AR · May 1, 2026
1,442
#13 · Gemma 4 31B
AR · May 1, 2026
1,432
#14 · Gemini 2.5 Pro
AR · May 1, 2026
1,429
#15 · Grok 4.20
AR · May 1, 2026
1,426
#16 · Claude Haiku 4.5
AR · May 1, 2026
1,424
#17 · Gemini 3 Flash
AR · May 1, 2026
1,421
#18 · GPT-5.1
AR · May 1, 2026
1,410
#19 · GPT-5.2
AR · May 1, 2026
1,406