Calibration
SWE-Rebench → SWE-bench Verified
SWE-bench Verified is estimated from SWE-Rebench with a Michaelis–Menten curve fitted on 8 models measured on both: y = 2.0000·x / (0.95233 + x), R² = 0.32, cross-validated error 4.8 pp. It is used for 0 estimates.
Fitted on
| Model | SWE-Rebench | SWE-bench Verified |
|---|---|---|
| Claude Opus 4.6 | 65.3% | 80.8% |
| Claude Sonnet 4.6 | 60.7% | 79.6% |
| GLM-4.7 | 58.7% | 73.8% |
| GLM-5 | 62.8% | 77.8% |
| GPT-5.3 Codex | 58.2% | 85.0% |
| Kimi K2.5 | 58.5% | 76.8% |
| Qwen3.5-27B | 58.9% | 72.4% |
| Qwen3.5-35B-A3B | 53.7% | 69.2% |