benchgap
Calibration

Toolathlon-Verified → AA-AnalystAgent

AA-AnalystAgent 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 / (0.82172^6.00 + x^6.00), R² = 0.52, cross-validated error 7.8 pp. It is used for 5 estimates.

Estimated modelToolathlon-VerifiedAA-AnalystAgentSource
DeepSeek V4 Pro 081374.1%42.0%estimated ± 7.8 pp, low confidence
Qwen3.8-Flash-Next73.5%40.6%estimated ± 7.8 pp, low confidence
Muse Spark Open72.8%39.1%estimated ± 7.8 pp, low confidence
Ornith-1.5-397B71.2%35.7%estimated ± 7.8 pp, low confidence
DeepSeek V4 Flash 073170.3%33.8%estimated ± 7.8 pp, low confidence

Fitted on

ModelToolathlon-VerifiedAA-AnalystAgent
Claude Fable 5.177.8%57.5%
Claude Opus 580.6%53.8%
Claude Opus 5.577.8%56.3%
Claude Sonnet 5.577.8%57.5%
GLM-5.373.0%38.8%
MiMo-V2.6-Pro76.9%37.5%
Step 5 Preview74.1%35.0%