CooperBench: leaderboard
Collaborative coding tasks from 12 open-source repositories in Python, TypeScript, Go and Rust: two agents run by the same model each implement one feature of a task pair in separate OpenHands containers and can message each other, and a task counts only if their merged patches pass both features' tests. CooperBench (Khatua et al., Stanford, 2026); percent of 652 tasks solved after the naive, union and learned-resolver merge stages, higher is better.
Metric: Cooperative task success (%; two agents that can message each other each implement one feature of a task pair in isolated workspaces, and the merged patch must pass both features' tests after the naive, union and learned-resolver merge stages; 652 tasks). Source: arxiv.org. Saturation forecast: Around February 2027. 5 models tracked.
Top models
| # | Model | Score | Overall rank |
|---|---|---|---|
| 1 | GPT-5 | 27.9 | #163 |
| 2 | Claude Sonnet 4.5 | 25.92 | #244 |
| 3 | MiniMax-M2 | 13.96 | #669 |
| 4 | Qwen 3 Coder 30B A3B Instruct | 13.34 | #973 |
| 5 | Qwen 3 30B A3B 2507 Instruct | 4.6 | #1078 |
No result here: #1 Claude Fable 5.2, #2 Gemini 4 Carbon, #5 Claude Opus 5.5.
Interactive version: theaggregate.ai/benchmark?slug=cooperbench · How It Works · Data refreshed daily, snapshot 2026-11-12.