benchgap
Calibration

Toolathlon-Verified → MazeBench

MazeBench is estimated from Toolathlon-Verified with a linear curve fitted on 10 models measured on both: y = 1.2724·x + -0.9156, R² = 0.61, cross-validated error 4.0 pp. It is used for 5 estimates.

Estimated modelToolathlon-VerifiedMazeBenchSource
DeepSeek V4 Pro 081374.1%2.7%estimated ± 4.0 pp, high confidence
Hy4 preview74.1%2.7%estimated ± 4.0 pp, high confidence
Step 5 Preview74.1%2.7%estimated ± 4.0 pp, high confidence
Qwen3.8-Flash-Next73.5%2.0%estimated ± 4.0 pp, high confidence
Ornith-1.5-397B71.2%0.0%estimated ± 4.0 pp, high confidence

Fitted on

ModelToolathlon-VerifiedMazeBench
Claude Fable 5.177.8%10.0%
Claude Opus 580.6%12.0%
Claude Opus 5.577.8%13.0%
Claude Sonnet 5.577.8%8.0%
DeepSeek V4 Flash 073170.3%0.0%
GLM-5.373.0%0.0%
GLM-5.3-Flash78.4%0.0%
Kimi K373.2%1.0%
Muse Spark Open72.8%0.0%
Qwen3.8 Max72.5%0.0%