benchgap
Calibration

NL2Repo → SWE-bench Pro

SWE-bench Pro is estimated from NL2Repo with a Hill curve fitted on 24 models measured on both: y = 0.3855 + (0.6329 − 0.3855)·x^6.00 / (0.33611^6.00 + x^6.00), R² = 0.72, cross-validated error 3.6 pp. It is used for 5 estimates.

Estimated modelNL2RepoSWE-bench ProSource
DeepSeek V4.1 Flash65.4%62.8%estimated ± 3.6 pp, medium confidence
GLM-5.358.0%62.4%estimated ± 3.6 pp, high confidence
GLM-5.3-Flash56.3%62.2%estimated ± 3.6 pp, high confidence
Seed 2.1 Pro47.0%60.4%estimated ± 3.6 pp, high confidence
Seed 2.1 Turbo43.7%59.0%estimated ± 3.6 pp, high confidence