Calibration

OSWorld 2.0 → MCP Atlas

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

MCP Atlas is estimated from OSWorld 2.0 with a Michaelis–Menten curve fitted on 8 models measured on both: y = 0.8455·x / (0.00856 + x), R² = 0.40, cross-validated error 10.4 pp. It is used for 1 estimate.

Estimated model	OSWorld 2.0	MCP Atlas	Source
Claude Opus 4.7	13.9%	79.6%	estimated ± 10.4 pp, low confidence