Arena Document

What does it measure?
Tests how well models read and analyze long, real user documents (PDFs) to answer questions, summarize, or extract information.
Who evaluates it, and how?
A user uploads a PDF and makes a request; two anonymous models respond and the user votes for the better one. The votes are combined into a score on the Elo scale, a rating built from head-to-head results.
Reading the score
Scores are Elo ratings from head-to-head human votes. Higher is better.
Top score1515
Models tested36
Scoring basisHigher is better
Last updated2026.11.09
Model release date

Which AI gives the document analysis answers people prefer most?

Arena Document

  1. 🥇AnthropicClaude Fable 5.11515
  2. 🥈AnthropicClaude Opus 5Reasoning effort: High1509
  3. 🥉AnthropicClaude Opus 4.6Reasoning effort: High1508
  4. #4AnthropicClaude Fable 51497
  5. #5AnthropicClaude Opus 4.7Reasoning effort: High1496
  6. #6OpenAIGPT-5.5Reasoning effort: High1484
  7. #7GoogleGemini 3.8 Flash1482
  8. #8AnthropicClaude Sonnet 4.61482
  9. #9OpenAIGPT-5.6 Sol1480
  10. #10GoogleGemini 3.7 Flash1476
  11. #11AnthropicClaude Opus 4.8Reasoning effort: High1475
  12. #12OpenAIGPT-5.41473
  13. #13MetaMuse Spark 1.31473
  14. #14OpenAIGPT-5.6 Terra1468
  15. #15OpenAIGPT-6 Astra1467
  16. #16AnthropicClaude Opus 4.51466
  17. #17MetaMuse Spark 1.11463
  18. #18GoogleGemini 3.5 FlashReasoning effort: High1462
  19. #19AnthropicClaude Sonnet 51460
  20. #20GoogleGemini 3.6 Flash1456
  21. #21OpenAIGPT-5.6 Luna1454
  22. #22SpaceXSIGrok 4.51451
  23. #23Moonshot AIKimi K2.61450
  24. #24AnthropicClaude Sonnet 4.51449
  25. #25MetaMuse Spark1444
  26. #26GoogleGemini 3.1 Pro1443
  27. #27SpaceXSIGrok 4.61438
  28. #28AlibabaQwen3.7 Plus1437
  29. #29MiniMaxMiniMax M31437
  30. #30GoogleGemma 4 31B1434
  31. #31Moonshot AIKimi K2.51431
  32. #32GoogleGemini 2.5 Pro1423
  33. #33AnthropicClaude Haiku 4.51420
  34. #34Z.aiGLM 5V Turbo1419
  35. #35SpaceXSIGrok 4.20 (Reasoning)1417
  36. #36GoogleGemini 3 Flash1414
Top 5 within the margin of error · effectively tied · Half of models ≤ 1461

The "harness" is the agent program the AI used to carry out the task. The same model can score very differently depending on its harness and reasoning effort.