Skip to content

ExtractBench: schema-guided enterprise document extraction (LlamaIndex leaderboard)

Scores schema-guided extraction from 370 enterprise documents and 4,869 pages on value F1, record completeness, word and page grounding and cost, split by short, medium and long documents.

24 models

Lower is better on this benchmark.

  1. 1Gemma 4 12BGoogle45.39
  2. 2Gemma 4 E2BGoogle51.85
  3. 3Gemma 4 E4BGoogle69.64
  4. 4Gemma 4 26B A4BGoogle70.27
  5. 5GPT-5.4 nanoOpenAI74.90
  6. 6Gemini 3.5 FlashGoogle79.84
  7. 7Gemini 3.8 FlashGoogle80.71
  8. 8GLM-5.3-FlashZ.AI80.75
  9. 9Kimi K3Moonshot AI83.17
  10. 10Claude Opus 5.5Anthropic86.06
  11. 11Claude Opus 4.8Anthropic87.09
  12. 12DeepSeek V4.1 Flash 1017DeepSeek87.11
  13. 13Qwen3.5-35B-A3BAlibaba88.00
  14. 14Qwen3.6-35B-A3BAlibaba88.11
  15. 15Qwen3.8-27BAlibaba89.75
  16. 16Qwen3.8-Flash-NextAlibaba89.88
  17. 17GPT-6 LunaOpenAI91.01
  18. 18GPT-5.6 LunaOpenAI91.16
  19. 19GPT-6 AstraOpenAI91.91
  20. 20GPT-5.6 TerraOpenAI92.26
  21. 21GPT-5.5OpenAI93.57
  22. 22GPT-5.6 SolOpenAI93.77
  23. 23GPT-6 SolOpenAI94.57
  24. 24Claude Sonnet 5.5Anthropic94.68
Lower is better. Bar colour: blue proprietary, orange open weights, green other.

About this benchmark

Metric
Order-insensitive value F1 overall and by document length; word and page grounding F1; cost and latency
Tasks
370 documents, 4,869 pages, 8 business domains, 67 document types, each with a JSON Schema
Authors
Boyang Zhang, Adrian Lyjak, Eli Stewart, Zhaoqi Li 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.