benchgap
Calibration

Toolathlon-Verified → DeepsecBench-AA

DeepsecBench-AA is estimated from Toolathlon-Verified with a Hill curve fitted on 8 models measured on both: y = 0.0000 + (0.6294 − 0.0000)·x^6.00 / (0.84970^6.00 + x^6.00), R² = 0.35, cross-validated error 5.2 pp. It is used for 9 estimates.

Estimated modelToolathlon-VerifiedDeepsecBench-AASource
Claude Sonnet 5.577.8%23.3%estimated ± 5.2 pp, low confidence
DeepSeek V4 Pro 081374.1%19.2%estimated ± 5.2 pp, low confidence
Hy4 preview74.1%19.2%estimated ± 5.2 pp, low confidence
MiMo-V2.6-Flash73.6%18.7%estimated ± 5.2 pp, low confidence
Qwen3.8-Flash-Next73.5%18.6%estimated ± 5.2 pp, low confidence
Muse Spark Open72.8%17.8%estimated ± 5.2 pp, low confidence
Qwen3.8 Max72.5%17.5%estimated ± 5.2 pp, low confidence
Ornith-1.5-397B71.2%16.2%estimated ± 5.2 pp, low confidence
DeepSeek V4 Flash 073170.3%15.3%estimated ± 5.2 pp, low confidence

Fitted on

ModelToolathlon-VerifiedDeepsecBench-AA
Claude Fable 5.177.8%23.5%
Claude Opus 580.6%25.9%
Claude Opus 5.577.8%26.9%
GLM-5.373.0%24.3%
GLM-5.3-Flash78.4%19.9%
Kimi K373.2%12.5%
MiMo-V2.6-Pro76.9%26.2%
Step 5 Preview74.1%15.8%