Skip to content

ParseBench: document parsing for AI agents (LlamaIndex leaderboard)

Scores document parsers and vision-language models on 2,078 human-verified enterprise pages across tables, charts, content faithfulness, semantic formatting and visual grounding, with the listed cost per page.

37 models

Higher is better.

  1. 1Claude Opus 5.5Anthropic79.85
  2. 2Claude Fable 5.1Anthropic78.92
  3. 3GPT-5.6 SolOpenAI75.35
  4. 4Gemini 3 FlashGoogle75.05
  5. 5Gemini 3.5 FlashGoogle74.02
  6. 6Gemini 3.8 FlashGoogle72.08
  7. 7Gemini 3.7 FlashGoogle71.27
  8. 8Qwen3.8-27BAlibaba70.79
  9. 9Claude Fable 5Anthropic70.78
  10. 10GLM-5.3-FlashZ.AI70.75
  11. 11GPT-6 AstraOpenAI70.67
  12. 12GPT-6.1 SolOpenAI70.48
  13. 13Claude Sonnet 5.5Anthropic70.18
  14. 14Gemini 3.6 FlashGoogle70.03
  15. 15GPT-6 SolOpenAI69.69
  16. 16Gemini 3.1 ProGoogle69.14
  17. 17GPT-5.6 TerraOpenAI68.39
  18. 18GPT-5.6 LunaOpenAI68.34
  19. 19GPT-5.5OpenAI67.76
  20. 20GPT-6 LunaOpenAI65.77
  21. 21DeepSeek V4.1 Flash 1017DeepSeek65.72
  22. 22Claude Opus 4.8Anthropic63.70
  23. 23Claude Opus 4.7Anthropic63.34
  24. 24Gemma 4 31BGoogle62.40
  25. 25GPT-5.4OpenAI62.23
  26. 26Claude Sonnet 5Anthropic62.13
  27. 27Gemini 3.5 Flash-LiteGoogle60.64
  28. 28Gemma 4 26B A4BGoogle58.50
  29. 29Gemini 3.1 Flash-LiteGoogle58.32
  30. 30Claude Haiku 5.5Anthropic57.40
  31. 31Claude Opus 4.6Anthropic54.07
  32. 32Nemotron 3 Nano Omni 30B A3BNVIDIA48.51
  33. 33Claude Haiku 4.5Anthropic45.17
  34. 34Qwen3.6-35B-A3BAlibaba44.10
  35. 35GPT-5.4 nanoOpenAI43.35
  36. 36Gemma 4 E4BGoogle40.50
  37. 37Amazon Nova 2 LiteAmazon40.43
Higher is better. Bar colour: blue proprietary, orange open weights, green other.

About this benchmark

Metric
Deterministic test rules per page; five dimension scores and their mean, in percent
Tasks
2,078 human-verified enterprise pages from insurance, finance and government documents
Authors
Boyang Zhang, Sebastián G. Acosta, Preston Carlson, Sacha Bron, Pierre-Loïc Doulcet, Daniel B. Ospina and Simon Suo (LlamaIndex) (2026)

Scores are as reported by model developers or evaluators and may use different settings (for example reasoning effort or tools). Compare with care.