Terminal-Bench-Science finds agents stuck at 30%

The important result is not the leaderboard winner but the low ceiling: real scientific work still breaks frontier agents routinely.

• –Tasks reflect research workflows rather than textbook questions, covering inference, simulation, optimization, theorem proving, image reconstruction, and scientific machine learning.
• –Three independent trials per task and reproducible, task-specific grading make the results more operationally meaningful than generic capability claims.
• –Claude Opus 5 leads GPT-5.6 Sol at 30.0% versus 22.4%; GPT-5.6 Sol matches Claude Fable 5's performance at substantially lower evaluation cost.
• –The open, continuously updated benchmark gives scientists a practical way to define what “AI for science” should actually accomplish.
• –With 70 tasks across five domains, future versions will need broader coverage to reduce variance and better represent specialized research.