benchgap
Calibration

Artificial Analysis Intelligence Index → Vals MMLU-Pro

Vals MMLU-Pro is estimated from Artificial Analysis Intelligence Index with a Michaelis–Menten + offset curve fitted on 47 models measured on both: y = 0.0000 + 0.9386·x / (0.02498 + x), R² = 0.49, cross-validated error 2.6 pp. It is used for 3 estimates.

Estimated modelArtificial Analysis Intelligence IndexVals MMLU-ProSource
Claude 4.1 Opus18.6%82.7%estimated ± 2.6 pp, medium confidence
o1-preview11.4%77.0%estimated ± 2.6 pp, low confidence
o3-pro21.9%84.2%estimated ± 2.6 pp, medium confidence