CharXiv
Can the model extract and reason over information in images or documents?
Vision14 ranked models17 reported values3 reportsHigher is better
Top rankings
One row per model, using its best reported score across effort settings.
| Rank | Model | Best score | Best reported setting |
|---|---|---|---|
| 1 | Claude Fable 5Anthropic | 88.9% | max effortMoonshot AI report1 value · 1 reportJul 17, 2026 · source |
| 2 | Kimi K3Moonshot AI | 84.8% | max effortMoonshot AI report1 value · 1 reportJul 17, 2026 · source |
| 3 | GPT-5.6 SolOpenAI | 84.6% | max effortMoonshot AI report1 value · 1 reportJul 17, 2026 · source |
| 4 | GPT-5.5OpenAI | 84.1% | xhigh effortMoonshot AI report1 value · 1 reportJul 17, 2026 · source |
| 5 | Qwen3.5-397B-A17BQwen | 80.8% | Reported configurationQwen report1 value · 1 reportApr 21, 2026 · source |
| 6 | Claude Opus 4.8Anthropic | 80.5% | max effortMoonshot AI report1 value · 1 reportJul 17, 2026 · source |
| 7 | Qwen3.5-27BQwen | 79.5% | Reported configurationQwen report2 values · 2 reportsApr 21, 2026 · source |
| 8 | Qwen3.6-27BQwen | 78.4% | Reported configurationQwen report1 value · 1 reportApr 21, 2026 · source |
| 9 | Qwen3.6-35B-A3BQwen | 78.0% | Reported configurationQwen report2 values · 2 reportsApr 21, 2026 · source |
| 10 | Qwen3.5-35B-A3BQwen | 77.5% | Reported configurationQwen report1 value · 1 reportApr 15, 2026 · source |
| 11 | Gemma4-26B-A4BGoogle | 69.0% | Reported configurationQwen report1 value · 1 reportApr 15, 2026 · source |
| 12 | Claude Opus 4.5Anthropic | 68.5% | Reported configurationQwen report1 value · 1 reportApr 21, 2026 · source |
Effort curve
Every sourced cost-linked effort value for this exact version. Lines connect complete sweeps only.
No cost-linked effort sweep for this version.
Definition and comparison boundarypublic methodology
Can the model extract and reason over information in images or documents?
Visual perception and multimodal reasoning under the reported protocol. Higher is better. The value is the percentage reported in this lab's table.
Harness, tools, prompts, attempt count, and benchmark version can materially change the result. Compare within one reporting context.
Method / source