benchgap
Calibration

Terminal-Bench-Science 0.1 → Terminal-Bench 4.0

Terminal-Bench 4.0 is estimated from Terminal-Bench-Science 0.1 with a Michaelis–Menten curve fitted on 11 models measured on both: y = 0.7823·x / (0.14226 + x), R² = 0.89, cross-validated error 6.9 pp. It is used for 0 estimates.

Fitted on

ModelTerminal-Bench-Science 0.1Terminal-Bench 4.0
Claude Fable 5.152.6%55.8%
Claude Fable 5.274.6%67.6%
Claude Mythos 5.275.3%70.5%
Claude Opus 5.558.7%66.4%
Claude Sonnet 5.559.9%70.6%
ERNIE 5.5 Flash Preview21.0%46.5%
GPT-6.1 Terra58.3%50.3%
GPT-6 Astra64.6%57.9%
Hy419.8%44.4%
Inkling 2 Mini1.4%5.6%
Kimi K3.1 Pro38.1%65.7%