CocoaBench: leaderboard
CocoaBench — domain reasoning benchmark from the every-eval-ever collection; accuracy on held-out tasks.
Metric: Accuracy. Source: cocoabench.github.io. Saturation forecast: Around March 2027. 10 models tracked.
Top models
| # | Model | Score | Overall rank |
|---|---|---|---|
| 1 | GPT-5.4 (High) | 45.1 | #79 (GPT-5.4) |
| 2 | Claude Sonnet 4.6 (High) | 34 | #90 (Claude Sonnet 4.6) |
| 3 | Gemini 3.1 Pro (Preview) | 30.7 | #58 |
| 4 | Gemini 3 Flash | 19.6 | #99 |
| 5 | Kimi K2.5 | 11.8 | #144 |
No result here: #3 Claude Opus 5.5, #6 GPT-6 Astra, #7 GPT-6.1 Sol.
Interactive version: theaggregate.ai/benchmark?slug=cocoabench · How It Works · Data refreshed daily, snapshot 2026-10-20.