Terminal Bench Science
A benchmark designed to measure the autonomous scientific research capabilities of AI models, evaluating their ability to independently complete tasks in scientific research domains.
Timeline (last 90 days)
GPT-6.1 Terra's score on the scientific programming test Terminal Bench Science exceeds that of the higher-positioned Soul and is less than 5 percentage points below Claude Opus 5.5, while its estimated cost is less than one-fifth of the latter's.
Gemini 4 Carbon ranks first on the Vending Bench long-horizon task benchmark.
In Terminal Bench Science at maximum reasoning intensity, Astra ranked first with 68.1%, and it is officially recommended for the most difficult scientific research tasks.
All Facts (3)
GPT-6.1 Terra's score on the scientific programming test Terminal Bench Science exceeds that of the higher-positioned Soul and is less than 5 percentage points below Claude Opus 5.5, while its estimated cost is less than one-fifth of the latter's.
50%UnverifiedGemini 4 Carbon ranks first on the Vending Bench long-horizon task benchmark.
50%UnverifiedIn Terminal Bench Science at maximum reasoning intensity, Astra ranked first with 68.1%, and it is officially recommended for the most difficult scientific research tasks.
50%