Calibration

LABBench2 → ARC-AGI-1

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

ARC-AGI-1 is estimated from LABBench2 with a Michaelis–Menten + offset curve fitted on 5 models measured on both: y = 0.3813 + 2.0000·x / (1.99797 + x), R² = 0.78, cross-validated error 0.7 pp. It is used for 2 estimates.

Estimated model	LABBench2	ARC-AGI-1	Source
Claude Sonnet 5	80.1%	95.4%	estimated ± 0.7 pp, low confidence
Gemini 4 Argon	88.8%	99.7%	estimated ± 0.7 pp, low confidence