PerceptionBench - Hallucination: leaderboard
PerceptionBench slice scoring perception-related hallucination — the share of questions answered without asserting content that is not in the image. The hardest slice on the board: no model clears 51%.
Metric: Perception Hallucination (%). Source: huggingface.co. Saturation forecast: Around 2029. 16 models tracked.
Top models
| # | Model | Score | Overall rank |
|---|---|---|---|
| 1 | Gemini 3.5 Flash | 50.6 | #66 |
| 2 | Seed 2.1 Pro | 49.8 | #73 |
| 3 | Claude Fable 5 | 45 | #20 |
| 4 | Grok 4.5 | 44.7 | #54 |
| 5 | Kimi K3 | 41.7 | #35 |
| 6 | Kimi K2.6 | 41 | #113 |
| 7 | Gemini 3.1 Pro (Preview) | 40.6 | #64 |
| 8 | Claude Opus 4.8 | 38.7 | #39 |
| 9 | GPT-5.5 | 34.7 | #33 |
| 10 | Gemma 4 31B | 32.1 | #212 |
| 11 | MiniMax-M3 | 29.9 | #151 |
| 12 | Qwen 3.7 Plus | 29.5 | #98 |
| 13 | GLM-5V Turbo | 28 | #172 |
| 14 | Qwen 3.5 397B A17B | 26.9 | #154 |
| 15 | GPT-5.6 Sol | 26.9 | #22 |
No result here: #3 Claude Opus 5.5, #6 GPT-6 Astra, #7 GPT-6.1 Sol.
Interactive version: theaggregate.ai/benchmark?slug=perceptionbench-hallucination · How It Works · Data refreshed daily, snapshot 2026-11-05.