benchgap
Calibration

Toolathlon → MCP Atlas

MCP Atlas is estimated from Toolathlon with a inverse Michaelis–Menten curve fitted on 17 models measured on both: y = 2.01861·(x − 0.0000) / (0.0000 + 2.0000 − x), R² = 0.58, cross-validated error 11.8 pp. It is used for 0 estimates.

Fitted on

ModelToolathlonMCP Atlas
Claude Opus 4.543.5%42.3%
Claude Opus 4.859.9%82.2%
DeepSeek V4 Flash 073147.8%69.0%
DeepSeek V4 Pro 081351.8%73.6%
Gemini 3.5 Flash56.5%83.6%
GLM-538.0%31.1%
GLM-5.248.2%76.8%
GPT-5.454.6%70.6%
GPT-5.4 mini42.9%57.7%
GPT-5.4 nano35.5%56.1%
GPT-5.555.6%75.3%
Kimi K2.650.0%55.9%
Kimi K2.527.8%29.5%
Muse Spark 1.175.6%88.1%
Qwen3.5 397B36.3%46.1%
Qwen3.6-35B-A3B26.9%62.8%
Qwen3.6 Plus39.8%48.2%