benchgap
Calibration

Vals SWE-bench → SciCode

SciCode is estimated from Vals SWE-bench with a Michaelis–Menten curve fitted on 8 models measured on both: y = 1.2863·x / (1.20621 + x), R² = 0.31, cross-validated error 4.5 pp. It is used for 2 estimates.

Estimated modelVals SWE-benchSciCodeSource
Composer 2.579.6%51.1%estimated ± 4.5 pp, medium confidence
Gemini 3.1 Flash-Lite62.8%44.0%estimated ± 4.5 pp, low confidence

Fitted on

ModelVals SWE-benchSciCode
Gemini 3.5 Flash78.8%53.1%
Grok 4.371.4%47.3%
Inkling-Small82.2%48.7%
Kimi K2.676.2%52.2%
Ling 3.0 Flash65.2%41.2%
Nemotron 3 Ultra69.0%44.6%
Qwen 3.6 Max (preview)72.8%47.0%
Qwen3.7 Max68.8%53.5%