PerceptionBench - Localization
Multimodal · 2026-07-16
The PerceptionBench localization category contains 330 questions that test whether a model can locate and anchor the correct region or object in an image.
Top models (higher is better)
| Model | Score |
|---|---|
| GPT-5.6 Sol | 76.7 |
| Claude Fable 5 | 70.4 |
| Kimi K3 | 70.3 |
| GPT-5.5 | 65.8 |
| Opus 4.8 | 58.8 |
| Gemini 3.1 Pro Preview | 52.7 |
| Qwen3.7-Plus | 52.7 |
| Gemini 3.5 Flash | 50.6 |
| Doubao Seed 2.1 Pro | 50.0 |
| Qwen3.5 397B A17B | 46.7 |
| Kimi K2.6 | 45.2 |
| Gemma 4 31B IT | 44.9 |
| GLM 5V Turbo | 43.9 |
| Grok 4.5 | 39.7 |
| MiniMax M3 | 33.3 |
| GLM-4.6V | 30.6 |