Roboflow Vision Evals — OCR (October)
Multimodal
Mean deterministic score over all questions in the redesigned OCR suite: single-value exact match, transcription character similarity, structured JSON field matching, and text localization requiring box IoU above 0.5 and exact text. Categories have unequal question counts; the overall score is not their unweighted mean. One preliminary run per effort tier; question count unreported.
Top models (higher is better)
| Model | Score |
|---|---|
| Claude Opus 5.5 | 79.9 |
| Gemini 3.7 Flash | 78.8 |
| Gemini 3.8 Flash | 78.7 |
| GPT-6 Astra | 77.8 |
| Gemini 3.6 Flash | 77.4 |
| Gemini 3.5 Flash | 76.7 |
| GPT-6.1 Sol | 75.8 |
| Claude Fable 5.1 | 75.3 |
| Claude Fable 5 | 74.2 |
| Gemini 3.1 Pro Preview | 73.1 |
| Gemini 3 Flash Preview | 73.1 |
| Gemini 3.5 Flash-Lite | 71.2 |
| Muse Spark 1.1 | 68.8 |
| GPT-6 Sol | 68.8 |
| Muse Spark 1.2 | 68.7 |
| Claude Opus 5 | 68.0 |
| Claude Sonnet 5.5 | 66.3 |
| Qwen3.7-Plus (source unspecified) | 65.5 |
| Claude Haiku 5.5 | 65.3 |
| GPT-5.6 Sol | 65.3 |
| Gemini 2.5 Pro | 65.2 |
| Grok 4.7 | 63.6 |
| Qwen3.8 Flash | 62.9 |
| GPT-6 Luna | 62.5 |
| Qwen3.7 Flash | 62.3 |
| Opus 4.8 | 62.1 |
| Kimi K3 | 61.8 |
| Qwen3.5-27B | 61.7 |
| Grok 4.6 | 61.2 |
| Muse Glimmer 30B | 60.7 |
| GPT-5.5 | 60.6 |
| Sonnet 5 | 59.5 |
| Grok 4.5 | 57.5 |
| GPT-5.6 Terra | 56.9 |
| Qwen3-VL-235B-A22B-Instruct | 56.4 |
| GLM 5V Turbo | 56.0 |
| GPT-5.4 Mini | 55.8 |
| GPT-5.6 Luna | 55.7 |
| GLM-5.3 Flash | 55.4 |
| Mistral Large 4 Preview | 54.5 |