benchgap
Calibration

HealthBench (length-adjusted) → ARC-AGI-1

ARC-AGI-1 is estimated from HealthBench (length-adjusted) with a Michaelis–Menten curve fitted on 6 models measured on both: y = 2.0000·x / (0.62861 + x), R² = 0.36, cross-validated error 4.4 pp. It is used for 1 estimate.

Estimated modelHealthBench (length-adjusted)ARC-AGI-1Source
GPT-6.1 Terra55.1%93.4%estimated ± 4.4 pp, low confidence

Fitted on

ModelHealthBench (length-adjusted)ARC-AGI-1
Claude Opus 557.8%97.5%
Claude Opus 5.560.6%97.5%
GPT-6.1 Sol58.5%96.5%
GPT-6 Astra58.3%97.5%
GPT-6 Luna54.5%86.7%
GPT-6 Sol53.2%95.5%