ParseBench: leaderboard
Document parsing benchmark from LlamaIndex over some 2,000 human verified pages from over 1,100 insurance, finance and government documents. Rule based checks score the parsed output on five dimensions, each averaged over its slice of the pages: table structure, chart data points, content faithfulness, semantic formatting and visual grounding. The overall score is their unweighted mean, higher is better.
Metric: Overall Score. Source: parsebench.ai. Saturation forecast: Around December 2027. 113 models tracked.
Top models
| # | Model | Score | Overall rank |
|---|---|---|---|
| 1 | Claude Opus 5.5 (High) | 79.85 | #3 (Claude Opus 5.5) |
| 2 | Claude Fable 5.1 | 78.92 | #9 |
| 3 | Claude Opus 5.5 | 78.01 | #3 |
| 4 | GPT-5.6 Sol (High) | 75.35 | #19 (GPT-5.6 Sol) |
| 5 | Gemini 3 Flash (High) | 75.05 | #99 (Gemini 3 Flash) |
| 6 | Gemini 3 Flash (Medium) | 75 | #99 (Gemini 3 Flash) |
| 7 | Claude Opus 5.5 (Low) | 74.47 | #3 (Claude Opus 5.5) |
| 8 | Gemini 3.5 Flash (High) | 74.02 | #60 (Gemini 3.5 Flash) |
| 9 | Gemini 3.8 Flash (High) | 72.08 | #21 (Gemini 3.8 Flash) |
| 10 | Gemini 3.7 Flash (High) | 71.27 | #23 (Gemini 3.7 Flash) |
| 11 | Gemini 3 Flash (Minimal) | 71.04 | #99 (Gemini 3 Flash) |
| 12 | Claude Fable 5 | 70.78 | #16 |
| 13 | Gemini 3.7 Flash (Medium) | 70.78 | #23 (Gemini 3.7 Flash) |
| 14 | GLM-5.3 Flash | 70.75 | #65 |
| 15 | Gemini 3.8 Flash (Medium) | 70.7 | #21 (Gemini 3.8 Flash) |
No result here: #8 Gemini 4 Argon, #10 Muse Wonder Full.
Interactive version: theaggregate.ai/benchmark?slug=parsebench · How It Works · Data refreshed daily, snapshot 2026-10-20.