CharXiv Reasoning (with tools)
Can the model extract and reason over information in images or documents?
Vision9 ranked models12 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 Opus 4.8Anthropic | 89.9% | max effortMeta report1 value · 1 reportJul 9, 2026 · source |
| 2 | Gemini 3.6 FlashGoogle | 89.4% | Reported configurationGoogle report2 values · 2 reportsAug 13, 2026 · source |
| 3 | Muse Spark 1.0Meta | 88.9% | Reported configurationMeta report1 value · 1 reportJul 9, 2026 · source |
| 4 | Gemini 3.7 FlashGoogle | 88.7% | Reported configurationGoogle report1 value · 1 reportAug 13, 2026 · source |
| 5 | Muse Spark 1.1Meta | 88.4% | xhigh effortMeta report1 value · 1 reportJul 9, 2026 · source |
| 6 | Claude Sonnet 5Anthropic | 88.3% | Reported configurationGoogle report2 values · 2 reportsAug 13, 2026 · source |
| 7 | Gemini 3.5 FlashGoogle | 84.9% | Reported configurationGoogle report1 value · 1 reportJul 21, 2026 · source |
| 8 | GPT-5.5OpenAI | 84.8% | xhigh effortMeta report1 value · 1 reportJul 9, 2026 · source |
| 9 | Gemini 3.1 ProGoogle | 83.2% | Reported configurationGoogle report2 values · 2 reportsJul 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.
With tools. Gemini's tool-enabled setup uses search and code execution. 1,000 questions; gpt-oss-120B high-reasoning judge; with code execution.
Method / source