benchgap
Calibration

Vals MMLU-Pro → HLE w/o tools

HLE w/o tools is estimated from Vals MMLU-Pro with a offset logistic curve fitted on 19 models measured on both: y = 0.2510 + (0.5738 − 0.2510) / (1 + exp(−64.54·(x − 0.8764))), R² = 0.86, cross-validated error 5.1 pp. It is used for 19 estimates.

Estimated modelVals MMLU-ProHLE w/o toolsSource
Claude Fable 591.5%54.9%estimated ± 5.1 pp, medium confidence
Claude Haiku 4.578.7%25.2%estimated ± 5.1 pp, medium confidence
Claude Opus 4.789.9%51.3%estimated ± 5.1 pp, medium confidence
Gemini 3.1 Flash-Lite86.2%34.2%estimated ± 5.1 pp, medium confidence
Gemini 3.5 Flash-Lite85.8%32.6%estimated ± 5.1 pp, medium confidence
Gemini 3.6 Flash89.3%49.1%estimated ± 5.1 pp, medium confidence
Gemini 3.7 Flash90.1%51.9%estimated ± 5.1 pp, medium confidence
Gemini 3 Flash88.6%46.1%estimated ± 5.1 pp, medium confidence
GLM-4.581.2%25.6%estimated ± 5.1 pp, medium confidence
GLM-4.682.2%26.0%estimated ± 5.1 pp, medium confidence
GLM-5.386.8%37.0%estimated ± 5.1 pp, medium confidence
GLM-5.3-Flash86.1%33.8%estimated ± 5.1 pp, medium confidence
Laguna M.168.8%25.1%estimated ± 5.1 pp, low confidence
Laguna XS.269.1%25.1%estimated ± 5.1 pp, low confidence
MiMo-V2.582.9%26.5%estimated ± 5.1 pp, medium confidence
MiniMax M384.2%28.3%estimated ± 5.1 pp, medium confidence
Mistral Medium 3.5 128B75.3%25.1%estimated ± 5.1 pp, low confidence
Muse Spark 1.288.3%44.6%estimated ± 5.1 pp, medium confidence
Qwen3.5 Flash84.1%28.1%estimated ± 5.1 pp, medium confidence