ZeroBench: leaderboard
ZeroBench holds 100 hand made visual reasoning questions, each needing several reasoning steps over one or more images, plus 334 easier subquestions covering single steps. Models give the final answer in curly braces, matched exactly against the key. This board is the official leaderboard's pass@5, the share of main questions solved in at least one of five samples run by the ZeroBench team, higher is better.
Metric: Score (%). Source: zerobench.github.io. Saturation forecast: Around July 2028. 74 models tracked.
Top models
| # | Model | Score | Overall rank |
|---|---|---|---|
| 1 | GPT-6 Astra (Max) | 52 | #6 (GPT-6 Astra) |
| 2 | Claude Opus 5.5 (Max) | 43 | #3 (Claude Opus 5.5) |
| 3 | GPT-6 Sol (Max) | 41 | #14 (GPT-6 Sol) |
| 4 | GPT-5.6 Sol (Max) | 30 | #19 (GPT-5.6 Sol) |
| 5 | Claude Opus 5 (Max) | 26 | #12 (Claude Opus 5) |
| 6 | Claude Fable 5 (Max) | 24 | #16 (Claude Fable 5) |
| 7 | GPT-5.5 (xHigh) | 23 | #31 (GPT-5.5) |
| 8 | GPT-5.4 (xHigh) | 23 | #79 (GPT-5.4) |
| 9 | Gemini 3.5 Flash | 22 | #60 |
| 10 | GPT-5.6 Luna (Max) | 21 | #68 (GPT-5.6 Luna) |
| 11 | Gemini 3 Pro | 20 | #80 |
| 12 | GPT-6 Luna (Max) | 20 | #56 (GPT-6 Luna) |
| 13 | Gemini 3.1 Pro (Preview) | 19 | #58 |
| 14 | GPT-5.6 Terra (Max) | 19 | #41 (GPT-5.6 Terra) |
| 15 | Claude Opus 4.8 | 18 | #37 |
No result here: #7 GPT-6.1 Sol, #8 Gemini 4 Argon, #9 Claude Fable 5.1.
Interactive version: theaggregate.ai/benchmark?slug=zerobench · How It Works · Data refreshed daily, snapshot 2026-10-20.