benchgap
Calibration

MCP Atlas → BrowseComp

BrowseComp is estimated from MCP Atlas with a Hill curve fitted on 23 models measured on both: y = 0.5716 + (1.1230 − 0.5716)·x^6.00 / (0.80751^6.00 + x^6.00), R² = 0.65, cross-validated error 8.4 pp. It is used for 3 estimates.

Estimated modelMCP AtlasBrowseCompSource
Kimi K2.7 Code76.0%79.8%estimated ± 8.4 pp, medium confidence
LLaDA2.2-flash46.2%59.0%estimated ± 8.4 pp, medium confidence
Solar Open 258.2%63.9%estimated ± 8.4 pp, medium confidence