benchgap
Calibration

Toolathlon-Verified → AA AutomationBench

AA AutomationBench is estimated from Toolathlon-Verified with a Hill curve fitted on 7 models measured on both: y = 0.0000 + (1.2000 − 0.0000)·x^6.00 / (0.75584^6.00 + x^6.00), R² = 0.46, cross-validated error 7.3 pp. It is used for 4 estimates.

Estimated modelToolathlon-VerifiedAA AutomationBenchSource
DeepSeek V4 Pro 081374.1%56.4%estimated ± 7.3 pp, low confidence
Hy4 preview74.1%56.4%estimated ± 7.3 pp, low confidence
Ornith-1.5-397B71.2%49.4%estimated ± 7.3 pp, low confidence
DeepSeek V4 Flash 073170.3%47.2%estimated ± 7.3 pp, low confidence

Fitted on

ModelToolathlon-VerifiedAA AutomationBench
Claude Fable 5.177.8%59.4%
Claude Opus 5.577.8%69.5%
Claude Sonnet 5.577.8%71.8%
GLM-5.373.0%62.2%
MiMo-V2.6-Pro76.9%58.6%
Muse Spark Open72.8%49.7%
Step 5 Preview74.1%51.0%