SWE-Shift (Clean)
Code · 2026-10-23
The matched no-change continuation of SWE-Shift: from the same saved history and workspace the agent continues without intervention, scored with the same strict criterion and family-cluster-scenario weighting over 120 scenarios and nine positions, in percent.
Top models (higher is better)
| Model | Score |
|---|---|
| GPT-6 Astra | 67.3 |
| Claude Fable 5.1 | 62.0 |
| Claude Opus 5 | 53.8 |
| GPT-5.6 Sol | 46.2 |
| Kimi K3 | 31.4 |
| DeepSeek-V4-Pro | 29.4 |
| DeepSeek-V4-Flash | 23.6 |
| Kimi K2.6 | 20.5 |
| MiniMax M3 | 19.8 |
| GLM-5.3 Flash | 19.5 |
| MiniMax M2.7 | 19.5 |
| Devstral 2 | 19.4 |
| Qwen3.8 27B | 13.2 |
| MiMo-V2.5-Pro | 12.0 |
| Qwen3 30B A3B Instruct 2507 | 11.2 |
| Qwen3-Coder-Next | 10.7 |