benchgap
Calibration

ARC-AGI-3 → BioBasicsBench

BioBasicsBench is estimated from ARC-AGI-3 with a linear curve fitted on 6 models measured on both: y = 0.4914·x + 0.3673, R² = 0.96, cross-validated error 3.5 pp. It is used for 3 estimates.

Estimated modelARC-AGI-3BioBasicsBenchSource
Claude Fable 5.265.7%69.0%estimated ± 3.5 pp, low confidence
Gemini 4 Carbon63.9%68.1%estimated ± 3.5 pp, low confidence
Grok 4.200.1%36.8%estimated ± 3.5 pp, low confidence

Fitted on

ModelARC-AGI-3BioBasicsBench
GPT-5.40.2%33.2%
GPT-5.50.4%35.4%
GPT-5.6 Sol7.8%45.6%
GPT-6.1 Sol52.7%62.6%
GPT-6 Astra62.7%67.0%
GPT-6 Luna0.1%37.6%