Roboflow Vision Evals — Counting, Strict (September)
Multimodal
Strict object-counting correctness requires exactly one integer matching ground truth. Reported beside judged accuracy under the September observed protocol.
Top models (higher is better)
| Model | Score |
|---|---|
| Gemini 3.5 Flash | 82.4 |
| Claude Opus 5.5 | 81.1 |
| GPT-6 Astra | 81.1 |
| Gemini 3.6 Flash | 80.2 |
| Gemini 3.7 Flash | 79.3 |
| Gemini 3.8 Flash | 79.3 |
| Muse Spark 1.1 | 76.6 |
| Muse Spark 1.2 | 76.6 |
| GPT-5.6 Sol | 76.1 |
| GPT-6 Sol | 76.1 |
| Muse Spark 1.3 | 74.8 |
| Gemini 3.1 Pro Preview | 71.6 |
| GPT-5.6 Luna | 70.7 |
| Claude Opus 5 | 70.3 |
| GPT-5.5 | 68.0 |
| Qwen3.8 Flash | 68.0 |
| Gemini 3 Flash Preview | 67.6 |
| Muse Glimmer 30B | 66.2 |
| GPT-5.6 Terra | 65.8 |
| GPT-6 Luna | 65.8 |
| Grok 4.6 | 65.8 |
| GPT-5.4 Mini | 64.9 |
| MiMo-V2.6-Flash | 64.4 |
| Claude Fable 5 | 63.5 |
| Grok 4.7 | 61.7 |
| Grok 4.5 | 59.5 |
| MiMo-V2.6-Pro | 58.1 |
| Sonnet 5 | 56.8 |
| GLM-5.3 Flash | 55.4 |
| Qwen3.5-27B | 54.0 |
| Opus 4.8 | 52.7 |
| Gemini 3.5 Flash-Lite | 52.7 |
| Gemini 2.5 Pro | 52.7 |
| Kimi K2.6 | 50.0 |
| Qwen3.7-Plus | 50.0 |
| Claude Fable 5.1 | 48.6 |
| GLM 5V Turbo | 48.6 |
| Qwen3-VL-235B-A22B-Instruct | 47.3 |
| Qwen3.7 Flash | 46.0 |
| Kimi K3 | 46.0 |