Benchmark
OCR Arena Win Rate
OCR benchmarks measure how well models read text in images and documents. Scores are percentages. Higher is better.
- Models measured
- 11
- Scores recorded
- 17
- Sources
- 1
- IQ weight
- 0.61
Score distribution
- 54.6%
- Mean
- 58.6%
- Median
- 16.5%
- Std dev
- 40%
- 10th percentile
- 70.3%
- 90th percentile
- 19.8%
- Min
This benchmark has a weight of 0.61 for chat models, measured over 9 models. The weight is how much it counts toward llminfo IQ. We measure it from the data, not from the raw scores.
Leaderboard
| # | Model | Score |
|---|---|---|
| 1 | Gemini 3 Flash· OCR Arena data | 77.1% |
| 2 | Gemini 2.5 Pro· OCR Arena data | 71.2% |
| 3 | Claude Opus 4.5medium· OCR Arena data | 70.3% |
| 4 | Claude Opus 4.6adaptive· OCR Arena data | 69.7% |
| 5 | Claude Opus 4.5low· OCR Arena data | 69.4% |
| 6 | Claude Opus 4.6medium· OCR Arena data | 67.9% |
| 7 | Claude Opus 4.6low· OCR Arena data | 63.8% |
| 8 | Gemini 2.5 Flash· OCR Arena data | 58.9% |
| 9 | GPT 5.1medium· OCR Arena data | 58.6% |
| 10 | GPT 5.1low· OCR Arena data | 54% |
| 11 | GPT 5.2none· OCR Arena data | 50.1% |
| 12 | GPT 5.2medium· OCR Arena data | 49.6% |
| 13 | GPT 5medium· OCR Arena data | 45.1% |
| 14 | GPT 5low· OCR Arena data | 42.7% |
| 15 | Qwen: Qwen3 VL 8B Thinking· OCR Arena data | 40% |
| 16 | DeepSeek-OCR· OCR Arena data | 20.5% |
| 17 | GLM OCR· OCR Arena data | 19.8% |