Roboflow Vision Evals (October 2026, five tasks)
Multimodal
Unweighted mean of five low-tier task headlines: detection mAP@50, the redesigned four-category OCR benchmark, and Gemini 3.5 Flash temperature-zero judged counting, identification and reasoning. Excludes models with missing OCR results. October marks the observed protocol change, not an upstream version number. Atomic sample counts unreported.
Top models (higher is better)
| Model | Score |
|---|---|
| GPT-6 Astra | 83.4 |
| Claude Opus 5.5 | 82.0 |
| Gemini 3.5 Flash | 81.8 |
| GPT-6.1 Sol | 81.7 |
| Gemini 3.7 Flash | 80.1 |
| Gemini 3.8 Flash | 79.8 |
| Claude Sonnet 5.5 | 77.6 |
| Gemini 3.6 Flash | 77.1 |
| GPT-6 Sol | 76.9 |
| Gemini 3.1 Pro Preview | 76.8 |
| Claude Fable 5.1 | 74.7 |
| Muse Spark 1.1 | 74.4 |
| Muse Spark 1.2 | 73.8 |
| GPT-5.6 Sol | 72.4 |
| Claude Fable 5 | 71.7 |
| Claude Opus 5 | 70.9 |
| GPT-6 Luna | 69.7 |
| Claude Haiku 5.5 | 69.6 |
| GPT-5.5 | 66.0 |
| GPT-5.6 Terra | 66.0 |
| Gemini 3 Flash Preview | 65.7 |
| GPT-5.6 Luna | 64.7 |
| Grok 4.7 | 63.5 |
| Gemini 3.5 Flash-Lite | 61.7 |
| Muse Glimmer 30B | 61.4 |
| Qwen3.8 Flash | 59.7 |
| Grok 4.6 | 59.1 |
| Qwen3.7-Plus (source unspecified) | 58.9 |
| Opus 4.8 | 58.3 |
| Gemini 2.5 Pro | 57.6 |
| Mistral Large 4 Preview | 56.8 |
| Kimi K3 | 56.0 |
| GLM-5.3 Flash | 55.8 |
| Grok 4.5 | 55.3 |
| Sonnet 5 | 55.3 |
| Qwen3-VL-235B-A22B-Instruct | 55.2 |
| GLM 5V Turbo | 54.6 |
| GPT-5.4 Mini | 53.2 |
| Kimi K2.6 | 52.8 |
| Qwen3.7 Flash | 52.4 |