Roboflow Vision Evals (September 2026 protocol)
Multimodal
Unweighted mean of six low-tier vision task headlines: detection mAP@50, OCR similarity, and Gemini 3.5 Flash temperature-zero judged accuracy for counting, identification, extraction and reasoning. Newer configurations average three runs; older configurations retain a single run. The September marker distinguishes this observed protocol from the earlier strict-match snapshot, not an upstream version number. The current pages do not state atomic sample counts.
Top models (higher is better)
| Model | Score |
|---|---|
| GPT-6 Astra | 86.6 |
| Gemini 3.5 Flash | 86.0 |
| Claude Opus 5.5 | 85.5 |
| Gemini 3.7 Flash | 85.2 |
| Gemini 3.8 Flash | 85.1 |
| Gemini 3.1 Pro Preview | 83.3 |
| Gemini 3.6 Flash | 83.0 |
| Claude Fable 5.1 | 81.3 |
| GPT-6 Sol | 80.7 |
| Muse Spark 1.1 | 80.5 |
| Muse Spark 1.2 | 80.5 |
| Muse Spark 1.3 | 79.8 |
| GPT-5.6 Sol | 79.0 |
| Claude Fable 5 | 78.7 |
| Claude Opus 5 | 78.3 |
| Gemini 3 Flash Preview | 74.9 |
| GPT-5.5 | 74.8 |
| GPT-5.6 Luna | 73.8 |
| GPT-5.6 Terra | 73.8 |
| Grok 4.7 | 71.9 |
| Muse Glimmer 30B | 70.8 |
| Gemini 3.5 Flash-Lite | 70.3 |
| Qwen3.8 Flash | 68.8 |
| Grok 4.6 | 68.7 |
| Opus 4.8 | 68.7 |
| GPT-6 Luna | 68.6 |
| Qwen3.7-Plus | 67.4 |
| Kimi K3 | 66.5 |
| Sonnet 5 | 66.4 |
| GLM-5.3 Flash | 66.3 |
| Gemini 2.5 Pro | 66.0 |
| Qwen3-VL-235B-A22B-Instruct | 65.9 |
| Grok 4.5 | 65.8 |
| GLM 5V Turbo | 65.3 |
| GPT-5.4 Mini | 64.7 |
| Qwen3.5-27B | 64.4 |
| Kimi K2.6 | 62.9 |
| MiMo-V2.6-Pro | 62.5 |
| Qwen3.7 Flash | 61.5 |
| MiMo-V2.6-Flash | 60.6 |