benchgap
Calibration

Toolathlon-Verified → GDPval-AA

GDPval-AA is estimated from Toolathlon-Verified with a inverse Michaelis–Menten curve fitted on 14 models measured on both: y = 0.43292·(x − 0.0000) / (0.0000 + 1.3230 − x), R² = 0.82, cross-validated error 5.6 pp. It is used for 7 estimates.

Estimated modelToolathlon-VerifiedGDPval-AASource
GLM-5.3-Flash78.4%63.0%estimated ± 5.6 pp, medium confidence
Qwen3.8 Max72.5%52.5%estimated ± 5.6 pp, medium confidence
Ornith-1.5-397B71.2%50.5%estimated ± 5.6 pp, medium confidence
dots3-note Preview55.6%31.4%estimated ± 5.6 pp, medium confidence
Laguna S 2.149.7%26.1%estimated ± 5.6 pp, low confidence
Ornith-1.5-35B-A3B48.7%25.2%estimated ± 5.6 pp, low confidence
Ornith-1.5-9B41.2%19.6%estimated ± 5.6 pp, low confidence

Fitted on

ModelToolathlon-VerifiedGDPval-AA
Claude Fable 5.177.8%62.7%
Claude Opus 580.6%61.1%
Claude Opus 5.577.8%68.3%
Claude Sonnet 5.577.8%66.9%
DeepSeek V4 Flash 073170.3%46.7%
DeepSeek V4 Pro 081374.1%54.5%
GLM-5.373.0%57.5%
Inkling-Small54.4%31.2%
Kimi K373.2%51.6%
MiMo-V2.6-Flash73.6%55.4%
MiMo-V2.6-Pro76.9%59.2%
Muse Spark Open72.8%45.1%
Qwen3.8-Flash-Next73.5%56.4%
Step 5 Preview74.1%54.3%