benchgap
Calibration

Vals LiveCodeBench → SWE-bench Verified

SWE-bench Verified is estimated from Vals LiveCodeBench with a offset logistic curve fitted on 21 models measured on both: y = 0.7626 + (0.9646 − 0.7626) / (1 + exp(−200.00·(x − 0.8777))), R² = 0.72, cross-validated error 4.4 pp. It is used for 8 estimates.

Estimated modelVals LiveCodeBenchSWE-bench VerifiedSource
Gemini 3.6 Flash88.1%89.6%estimated ± 4.4 pp, high confidence
Gemini 3.8 Flash89.5%95.8%estimated ± 4.4 pp, high confidence
GLM-4.567.4%76.3%estimated ± 4.4 pp, high confidence
Grok 4.384.5%76.3%estimated ± 4.4 pp, high confidence
Grok 4.688.2%90.5%estimated ± 4.4 pp, high confidence
Kimi K2.7 Code82.1%76.3%estimated ± 4.4 pp, high confidence
Kimi K387.2%81.2%estimated ± 4.4 pp, high confidence
Qwen3.5 Flash83.3%76.3%estimated ± 4.4 pp, high confidence