benchgap
Calibration

SWE Multilingual → MILU

MILU is estimated from SWE Multilingual with a Hill curve fitted on 5 models measured on both: y = 0.8880 + (1.2000 − 0.8880)·x^6.00 / (1.26794^6.00 + x^6.00), R² = 0.62, cross-validated error 0.9 pp. It is used for 4 estimates.

Estimated modelSWE MultilingualMILUSource
Claude Mythos 592.2%92.8%estimated ± 0.9 pp, medium confidence
Ling 3.1 Flash88.5%92.0%estimated ± 0.9 pp, low confidence
Claude Opus 4.884.4%91.3%estimated ± 0.9 pp, low confidence
Hy4 preview82.9%91.1%estimated ± 0.9 pp, low confidence

Fitted on

ModelSWE MultilingualMILU
Claude Fable 5.189.1%93.0%
Claude Fable 5.296.6%94.1%
Claude Opus 589.5%92.1%
Claude Opus 5.593.9%93.1%
Claude Sonnet 5.590.3%91.6%