benchgap
Calibration

SWE-bench Pro → SciCode

SciCode is estimated from SWE-bench Pro with a linear curve fitted on 19 models measured on both: y = 0.4180·x + 0.2421, R² = 0.61, cross-validated error 5.6 pp. It is used for 3 estimates.

Estimated modelSWE-bench ProSciCodeSource
Hy4 preview65.7%51.7%estimated ± 5.6 pp, medium confidence
Atria Dawn Preview59.6%49.1%estimated ± 5.6 pp, medium confidence
Laguna S 2.159.4%49.0%estimated ± 5.6 pp, medium confidence

Fitted on

ModelSWE-bench ProSciCode
Gemini 3.5 Flash55.1%53.1%
Granite 4.2 30B33.3%38.8%
Granite 4.2 8B19.1%36.1%
Inkling-Small55.9%48.7%
Kimi K2.658.6%52.2%
Kimi K2.550.7%48.7%
Ling 3.0 Flash56.6%41.2%
MiniCPM5-2B14.4%26.3%
Muse Glimmer 30B51.2%43.6%
Qwen 3.6 Max (preview)57.3%47.0%
Qwen3.7 Max60.6%53.5%
Qwen3.7 Plus57.6%51.3%
Beam 1.0 501B-A23B66.1%49.2%
Beam 1.0 501B-A23B FP865.9%48.3%
Beam 1.0 501B-A23B NVFP465.6%46.9%
Sage-A1.561.2%44.3%
Sakana Fugu59.0%60.1%
Sakana Fugu-Ultra73.7%58.7%
Spark-X2.6-4B54.3%36.8%