benchgap
Calibration

WideResearch → Toolathlon-Verified

Toolathlon-Verified is estimated from WideResearch with a inverse Michaelis–Menten curve fitted on 6 models measured on both: y = 0.64122·(x − 0.0000) / (0.0000 + 1.5727 − x), R² = 0.89, cross-validated error 5.5 pp. It is used for 3 estimates.

Estimated modelWideResearchToolathlon-VerifiedSource
Claude Opus 4.576.4%60.6%estimated ± 5.5 pp, low confidence
Qwen3.5 397B74.0%57.0%estimated ± 5.5 pp, low confidence
GLM-569.8%51.2%estimated ± 5.5 pp, low confidence

Fitted on

ModelWideResearchToolathlon-Verified
dots3-note Preview78.9%55.6%
Hy4 preview83.9%74.1%
Ornith-1.5-35B-A3B67.8%48.7%
Ornith-1.5-397B80.8%71.2%
Ornith-1.5-9B59.5%41.2%
Qwen3.8 Max81.9%72.5%