benchgap
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

ModelSWE-RebenchSWE-bench Verified
Claude Opus 4.665.3%80.8%
Claude Sonnet 4.660.7%79.6%
GLM-4.758.7%73.8%
GLM-562.8%77.8%
GPT-5.3 Codex58.2%85.0%
Kimi K2.558.5%76.8%
Qwen3.5-27B58.9%72.4%
Qwen3.5-35B-A3B53.7%69.2%