DeepSWE v1
Code · 2026-05-26
DeepSWE v1 evaluates frontier coding agents on 113 original, long-horizon software-engineering tasks across 91 repositories. Pass@1 is the pooled pass rate over scored rollout attempts; context-window failures and agent timeouts count as failures, while provider, verifier, and network errors are excluded.
Top models (higher is better)
| Model | Score |
|---|---|
| GPT-5.5 | 70.0 |
| Opus 4.8 | 58.2 |
| GPT-5.4 | 55.5 |
| Opus 4.7 | 54.2 |
| GLM-5.2 | 41.5 |
| Sonnet 4.6 | 31.8 |
| Gemini 3.5 Flash | 28.3 |
| Opus 4.6 | 27.6 |
| GPT-5.4 Mini | 24.3 |
| Kimi K2.6 | 23.9 |
| MiniMax M3 | 20.4 |
| MiMo-V2.5-Pro | 19.5 |
| Qwen3.7-Max | 17.7 |
| GLM-5.1 | 17.5 |
| Grok Build 0.1 | 13.1 |
| Gemini 3.1 Pro Preview | 9.7 |
| DeepSeek-V4-Pro | 7.5 |
| Gemini 3 Flash Preview | 5.1 |
| Qwen3.6 Plus (2026-04-02) | 2.7 |
| Haiku 4.5 | 0.2 |
| MiniMax M2.7 | 0.2 |