Calibration

MMAnswerBench → FrontierMath v2 (Tier 4)

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

FrontierMath v2 (Tier 4) is estimated from MMAnswerBench with a linear curve fitted on 6 models measured on both: y = 2.7467·x + -2.2211, R² = 0.62, cross-validated error 3.7 pp. It is used for 4 estimates.

Estimated model	MMAnswerBench	FrontierMath v2 (Tier 4)	Source
GLM-5.2	91.0%	27.8%	estimated ± 3.7 pp, low confidence
Qwen3.5 397B	80.9%	0.1%	estimated ± 3.7 pp, low confidence
Qwen3.6-27B	80.8%	0.0%	estimated ± 3.7 pp, low confidence
Qwen3.6-35B-A3B	78.9%	0.0%	estimated ± 3.7 pp, low confidence