benchgap
Calibration

Toolathlon → Gert Labs

Gert Labs is estimated from Toolathlon with a Hill curve fitted on 13 models measured on both: y = 0.4542 + (1.2000 − 0.4542)·x^6.00 / (0.65907^6.00 + x^6.00), R² = 0.67, cross-validated error 7.2 pp. It is used for 3 estimates.

Estimated modelToolathlonGert LabsSource
GLM-5.248.2%55.3%estimated ± 7.2 pp, medium confidence
GPT-5.4 mini42.9%50.7%estimated ± 7.2 pp, medium confidence
GPT-5.4 nano35.5%47.2%estimated ± 7.2 pp, medium confidence

Fitted on

ModelToolathlonGert Labs
Claude Opus 4.543.5%64.2%
Claude Opus 4.859.9%73.0%
Gemini 3.5 Flash56.5%61.9%
GLM-538.0%51.0%
GPT-5.454.6%64.9%
GPT-5.555.6%72.9%
Kimi K2.650.0%56.8%
Kimi K2.527.8%45.9%
MiniMax M2.746.3%40.4%
Qwen3.5 397B36.3%46.8%
Qwen3.6-35B-A3B26.9%42.7%
Qwen3.6 Plus39.8%50.6%
Step 3.7 Flash49.5%51.6%