Benchmarks / Coding
DeepSWE v1
Original DeepSWE leaderboard: pass@1 on 113 long-horizon SWE tasks with mini-swe-agent. Used as fallback for models not yet run on v1.1.
DeepSWE v1
Higher is better · best effort per model
full table below
| # | Model | Effort | Score |
|---|---|---|---|
| 1 | GPT-5.5 | xhigh | 70.0% |
| 2 | Opus 4.8 | max | 58.2% |
| 3 | GPT-5.4 | xhigh | 55.5% |
| 4 | Opus 4.7 | max | 54.2% |
| 5 | GLM 5.2 | max | 41.5% |
| 6 | Sonnet 4.6 | high | 31.8% |
| 7 | Gemini 3.5 Flash | medium | 28.3% |
| 8 | Opus 4.6 | max | 27.6% |
| 9 | GPT-5.4 Mini | xhigh | 24.3% |
| 10 | Kimi K2.6 | max | 23.9% |
| 11 | MiniMax M3 | max | 20.4% |
| 12 | MiMo V2.5 Pro | xhigh | 19.5% |
| 13 | Qwen 3.7 Max | max | 17.7% |
| 14 | GLM 5.1 | max | 17.5% |
| 15 | Grok Build 0.1 | – | 13.1% |
| 16 | Gemini 3.1 Pro | high | 9.7% |
| 17 | DeepSeek V4 Pro | max | 7.5% |
| 18 | Gemini 3 Flash | high | 5.1% |
| 19 | Qwen 3.6 Plus | max | 2.7% |
| 20 | Haiku 4.5 | xhigh | 0.2% |
| 21 | MiniMax M2.7 | high | 0.2% |