benchgap
Calibration

Vals MMLU-Pro → GPQA

GPQA is estimated from Vals MMLU-Pro with a linear curve fitted on 24 models measured on both: y = 0.9549·x + 0.0770, R² = 0.77, cross-validated error 1.6 pp. It is used for 22 estimates.

Estimated modelVals MMLU-ProGPQASource
Claude Fable 591.5%95.1%estimated ± 1.6 pp, medium confidence
Claude Fable 5.192.4%95.9%estimated ± 1.6 pp, medium confidence
Claude Haiku 4.578.7%82.9%estimated ± 1.6 pp, high confidence
Claude Opus 4.789.9%93.5%estimated ± 1.6 pp, medium confidence
Claude Sonnet 587.5%91.3%estimated ± 1.6 pp, high confidence
Gemini 3.1 Flash-Lite86.2%90.0%estimated ± 1.6 pp, high confidence
Gemini 3.5 Flash-Lite85.8%89.6%estimated ± 1.6 pp, high confidence
Gemini 3.6 Flash89.3%93.0%estimated ± 1.6 pp, high confidence
Gemini 3.7 Flash90.1%93.7%estimated ± 1.6 pp, medium confidence
Gemini 3 Flash88.6%92.3%estimated ± 1.6 pp, high confidence
GLM-4.581.2%85.2%estimated ± 1.6 pp, high confidence
GLM-4.682.2%86.2%estimated ± 1.6 pp, high confidence
GLM-5.386.8%90.6%estimated ± 1.6 pp, high confidence
GLM-5.3-Flash86.1%89.9%estimated ± 1.6 pp, high confidence
Laguna M.168.8%73.4%estimated ± 1.6 pp, medium confidence
Laguna XS.269.1%73.7%estimated ± 1.6 pp, medium confidence
MiMo-V2.582.9%86.9%estimated ± 1.6 pp, high confidence
MiMo-V2.5-Pro84.6%88.5%estimated ± 1.6 pp, high confidence
MiniMax M384.2%88.1%estimated ± 1.6 pp, high confidence
Mistral Medium 3.5 128B75.3%79.6%estimated ± 1.6 pp, medium confidence
Muse Spark 1.288.3%92.0%estimated ± 1.6 pp, high confidence
Qwen3.5 Flash84.1%88.0%estimated ± 1.6 pp, high confidence