PerceptionBench: leaderboard
Moonshot AI benchmark for atomic visual perception in multimodal models: 3,000 verified open-ended questions with short, uniquely determined answers across ten perceptual capabilities, graded by an LLM judge against the reference. Sixteen frontier models are run with unified prompts at their highest reasoning budget, and none reaches 60%.
Metric: Accuracy (%). Source: huggingface.co. Saturation forecast: Around October 2027. 16 models tracked.
Top models
| # | Model | Score | Overall rank |
|---|---|---|---|
| 1 | GPT-5.6 Sol | 59.7 | #22 |
| 2 | Kimi K3 | 58.5 | #35 |
| 3 | Claude Fable 5 | 57.2 | #20 |
| 4 | Gemini 3.1 Pro (Preview) | 56.2 | #64 |
| 5 | GPT-5.5 | 55.8 | #33 |
| 6 | Seed 2.1 Pro | 55 | #73 |
| 7 | Gemini 3.5 Flash | 52 | #66 |
| 8 | Qwen 3.7 Plus | 51.1 | #98 |
| 9 | Qwen 3.5 397B A17B | 47.5 | #154 |
| 10 | Claude Opus 4.8 | 47.2 | #39 |
| 11 | Kimi K2.6 | 42.6 | #113 |
| 12 | Grok 4.5 | 41 | #54 |
| 13 | Gemma 4 31B | 40.7 | #212 |
| 14 | GLM-5V Turbo | 39.6 | #172 |
| 15 | MiniMax-M3 | 33.1 | #151 |
No result here: #3 Claude Opus 5.5, #6 GPT-6 Astra, #7 GPT-6.1 Sol.
Interactive version: theaggregate.ai/benchmark?slug=perceptionbench · How It Works · Data refreshed daily, snapshot 2026-11-05.