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 Hill curve fitted on 21 models measured on both: y = 0.0010 + (1.2000 − 0.0010)·x^0.52 / (0.65971^0.52 + x^0.52), R² = 0.86, cross-validated error 8.2 pp. It is used for 2 estimates.

Estimated modelTerminal-Bench Science 0.1Terminal-Bench 4.0Source
Claude Opus 5.5 (medium with fallback)43.3%53.6%estimated ± 8.2 pp, medium confidence
Claude Sonnet 5.5 (high with fallback)34.3%50.0%estimated ± 8.2 pp, medium confidence