Roboflow Vision Evals — Reasoning, Strict (September)
Multimodal
Normalized exact-match visual reasoning accuracy, reported beside judged accuracy under the September observed protocol.
Top models (higher is better)
| Model | Score |
|---|---|
| GPT-6 Astra | 89.6 |
| GPT-6.1 Sol | 87.0 |
| Claude Opus 5.5 | 85.9 |
| Gemini 3.8 Flash | 84.5 |
| Gemini 3.7 Flash | 81.2 |
| Gemini 3.5 Flash | 81.0 |
| Gemini 3.6 Flash | 81.0 |
| GPT-6 Sol | 77.3 |
| Muse Spark 1.1 | 74.8 |
| Gemini 3.1 Pro Preview | 74.8 |
| Muse Spark 1.2 | 74.6 |
| Gemini 3 Flash Preview | 74.2 |
| Kimi K3 | 74.2 |
| Muse Spark 1.3 | 73.3 |
| Claude Sonnet 5.5 | 72.8 |
| Claude Haiku 5.5 | 72.8 |
| GPT-5.5 | 72.2 |
| GPT-5.6 Sol | 71.7 |
| GPT-6 Luna | 70.2 |
| Qwen3.8 Flash | 69.5 |
| Claude Opus 5 | 68.9 |
| Gemini 3.5 Flash-Lite | 68.9 |
| Grok 4.7 | 66.7 |
| Qwen3.7-Plus (source unspecified) | 66.2 |
| Claude Fable 5 | 66.2 |
| GPT-5.6 Luna | 65.6 |
| GPT-5.6 Terra | 65.3 |
| GPT-5.4 Mini | 64.0 |
| Grok 4.6 | 63.8 |
| Muse Glimmer 30B | 62.9 |
| Gemini 2.5 Pro | 62.3 |
| Qwen3.5-27B | 61.6 |
| Qwen3.7 Flash | 60.9 |
| Grok 4.5 | 59.8 |
| Kimi K2.6 | 57.6 |
| Mistral Large 4 Preview | 57.4 |
| MiMo-V2.6-Flash | 55.9 |
| MiMo-V2.6-Pro | 54.1 |
| GLM-5.3 Flash | 51.7 |
| GLM 5V Turbo | 49.7 |