Calibration

ARC-AGI-2 → ARC-AGI-3

Name: benchgap: LLM benchmark scores, measured and estimated
Creator: benchgap

ARC-AGI-3 is estimated from ARC-AGI-2 with a inverse Michaelis–Menten curve fitted on 17 models measured on both: y = 0.11942·(x − 0.8284) / (0.8284 + 0.1439 − x), R² = 0.89, cross-validated error 13.8 pp. It is used for 2 estimates.

Estimated model	ARC-AGI-2	ARC-AGI-3	Source
Claude Sonnet 4.5 Thinking	13.6%	0.0%	estimated ± 13.8 pp, low confidence
Gemini 3 Pro Deep Think	45.1%	0.0%	estimated ± 13.8 pp, low confidence