ZeroBench: leaderboard
"Impossible" visual benchmark for multimodal models. Best pass@5 is only 19%, 5/5 reliability just 6%. Exposes limits of current vision-language reasoning.
Metric: Score (%). Source: zerobench.github.io. Status: years away from saturation. 70 models tracked.
Top models
| # | Model | Score |
|---|---|---|
| 1 | GPT-5.6 Sol (Max) | 30 |
| 2 | Claude Opus 5 (Max) | 26 |
| 3 | Claude Fable 5 (Max) | 24 |
| 4 | GPT-5.5 (xHigh) | 23 |
| 5 | GPT-5.4 (xHigh) | 23 |
| 6 | Gemini 3.5 Flash | 22 |
| 7 | GPT-5.6 Luna (Max) | 21 |
| 8 | Gemini 3 Pro | 20 |
| 9 | Gemini 3.1 Pro (Preview) | 19 |
| 10 | GPT-5.6 Terra (Max) | 19 |
| 11 | Claude Opus 4.8 | 18 |
| 12 | GPT-5.2 (Medium) | 18 |
| 13 | Gemini 3.6 Flash (High) | 18 |
| 14 | Grok 4.6 (xHigh) | 17 |
| 15 | Gemini 3 Flash | 15 |
Interactive version: theaggregate.ai/benchmark?slug=zerobench · How It Works · Data refreshed daily, snapshot 2026-09-05.