Calibration

HLE w/ tools → MCP Atlas

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

MCP Atlas is estimated from HLE w/ tools with a Michaelis–Menten curve fitted on 8 models measured on both: y = 1.2688·x / (0.34234 + x), R² = 0.73, cross-validated error 7.3 pp. It is used for 4 estimates.

Estimated model	HLE w/ tools	MCP Atlas	Source
Agents-A1	47.6%	73.8%	estimated ± 7.3 pp, medium confidence
Claude Sonnet 5	57.4%	79.5%	estimated ± 7.3 pp, medium confidence
Qwen3.8 Max	56.2%	78.9%	estimated ± 7.3 pp, medium confidence
Step 3.7 Flash	47.2%	73.5%	estimated ± 7.3 pp, medium confidence