Roboflow Vision Evals — Extraction, Strict (September)
Multimodal
Normalized exact-match image data-extraction accuracy, reported beside judged accuracy under the September observed protocol.
Top models (higher is better)
| Model | Score |
|---|---|
| Gemini 3 Flash Preview | 95.9 |
| Gemini 3.7 Flash | 95.5 |
| Gemini 3.8 Flash | 95.2 |
| Gemini 3.1 Pro Preview | 94.8 |
| Gemini 3.6 Flash | 94.5 |
| Gemini 3.5 Flash | 94.5 |
| Claude Fable 5.1 | 92.1 |
| Claude Opus 5.5 | 92.1 |
| Claude Sonnet 5.5 | 91.8 |
| Claude Fable 5 | 91.8 |
| Gemini 3.5 Flash-Lite | 90.7 |
| GPT-6 Astra | 90.0 |
| GPT-6.1 Sol | 89.0 |
| Sonnet 5 | 88.7 |
| Claude Opus 5 | 88.7 |
| Muse Spark 1.2 | 88.7 |
| Muse Spark 1.1 | 88.3 |
| Opus 4.8 | 87.6 |
| Muse Spark 1.3 | 87.6 |
| Grok 4.7 | 87.6 |
| Qwen3-VL-235B-A22B-Instruct | 86.6 |
| GPT-5.6 Sol | 85.9 |
| GPT-5.5 | 85.9 |
| Muse Glimmer 30B | 85.6 |
| Grok 4.6 | 84.5 |
| Kimi K3 | 84.5 |
| Qwen3.8 Flash | 84.2 |
| GPT-6 Luna | 83.8 |
| Gemini 2.5 Pro | 83.5 |
| GPT-5.4 Mini | 83.2 |
| GPT-6 Sol | 83.2 |
| Grok 4.5 | 82.8 |
| Qwen3.7-Plus (source unspecified) | 82.5 |
| GLM-5.3 Flash | 81.4 |
| MiMo-V2.6-Flash | 81.4 |
| Mistral Large 4 Preview | 81.4 |
| GPT-5.6 Luna | 80.8 |
| MiMo-V2.6-Pro | 79.7 |
| GLM 5V Turbo | 79.4 |
| GPT-5.6 Terra | 79.4 |