benchgap
Calibration

Vals MMLU-Pro → HLE

HLE is estimated from Vals MMLU-Pro with a inverse Michaelis–Menten curve fitted on 29 models measured on both: y = 0.08531·(x − 0.0000) / (0.0000 + 1.0439 − x), R² = 0.48, cross-validated error 9.4 pp. It is used for 16 estimates.

Estimated modelVals MMLU-ProHLESource
Claude Opus 4.789.9%52.9%estimated ± 9.4 pp, low confidence
Gemini 3 Flash88.6%47.9%estimated ± 9.4 pp, low confidence
Muse Spark 1.288.3%46.8%estimated ± 9.4 pp, low confidence
GLM-5.386.8%42.1%estimated ± 9.4 pp, low confidence
Gemini 3.1 Flash-Lite86.2%40.4%estimated ± 9.4 pp, low confidence
GLM-5.3-Flash86.1%40.1%estimated ± 9.4 pp, low confidence
Gemini 3.5 Flash-Lite85.8%39.4%estimated ± 9.4 pp, low confidence
MiniMax M384.2%35.6%estimated ± 9.4 pp, low confidence
Qwen3.5 Flash84.1%35.4%estimated ± 9.4 pp, low confidence
MiMo-V2.582.9%32.9%estimated ± 9.4 pp, low confidence
GLM-4.682.2%31.6%estimated ± 9.4 pp, low confidence
GLM-4.581.2%29.9%estimated ± 9.4 pp, low confidence
Claude Haiku 4.578.7%26.1%estimated ± 9.4 pp, low confidence
Mistral Medium 3.5 128B75.3%22.1%estimated ± 9.4 pp, low confidence
Laguna XS.269.1%16.7%estimated ± 9.4 pp, low confidence
Laguna M.168.8%16.5%estimated ± 9.4 pp, low confidence