Boyang Zhang et al.

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.

In index
Multimodal
Best in multimodal
Source ↗

Models39

Top result78.8%Claude Opus 5.5

Top-3 spread4.4 pts

Year2026

Result and price

33 of 39 results

Results

39 results

#ModelResult

01Claude Opus 5.578.8%

02Claude Fable 5.177.8%

03GPT-5.6 Sol74.4%

04Gemini 3 FlashAbove fit74.1%

05GPT-6.1 Terra73.6%

06Gemini 3.5 Flash73.2%

07Gemini 3.8 Flash71.2%

08GLM-5.3-Flash70.5%

09Gemini 3.7 Flash70.4%

10GPT-6 Astra69.8%

10Claude Fable 569.8%

12GPT-6.1 Sol69.6%

13Claude Sonnet 5.569.2%

14Gemini 3.6 Flash69.1%

15Qwen3.8-27B69.0%

16GPT-6 Sol68.8%

17Gemini 3.1 Pro67.3%

18GPT-5.6 Terra67.1%

19GPT-5.6 Luna66.9%

20GPT-5.566.4%

21DeepSeek V4.1 Flash 101764.4%

22GPT-6 Luna64.1%

23Claude Opus 4.763.3%

24Gemma 4 31B62.4%

25Claude Opus 4.862.3%

26GPT-5.461.7%

27Claude Sonnet 561.2%

28Gemini 3.5 Flash-Lite59.5%

29Gemma 4 26B A4B58.5%

30Gemini 3.1 Flash-Lite56.7%

31Claude Haiku 5.556.0%

32Claude Opus 4.653.8%

33Claude Haiku 4.5 Thinking51.2%

34Claude Haiku 4.549.3%

35Nemotron 3 Nano Omni 30B A3B48.5%

36GPT-5.4 nano42.2%

37Gemma 4 E4B41.1%

38Amazon Nova 2 Lite40.4%

39Qwen3.6-35B-A3B37.4%