What is Terminal-Bench-Science?

Terminal-Bench-Science is a benchmark that gives an AI agent a real scientific research workflow, such as fitting a model, running a simulation, or proving a theorem, and passes it only if the task’s deterministic tests confirm the result. Version 0.1, released August 27, 2026, has 70 tasks across the life, physical, Earth, mathematical, and engineering sciences.

It’s led by researchers at Stanford University and built by the team behind Terminal-Bench, with every task written by a domain expert in its own field. The Terminal-Bench-Science 0.1 announcement reports 920 proposals, 386 pull requests, and 70 merged tasks, each cleared by domain review, technical review, and a final bar-raiser check. As of October 18, 2026, 0.1 is the only release, and the project’s contribution page is collecting tasks for version 1.0 until December 15, 2026; the team says later releases will add tasks and retire the ones agents saturate.

In the 0.1 announcement, the top result was Anthropic’s Claude Opus 5 running in Claude Code at 30.0%, averaged over three trials per task. Each entry is a model-and-harness pair, and a vendor’s own harness isn’t always the best one. Anthropic’s September 22, 2026 Claude Opus 5.5 announcement later reported 58.7% for Opus 5.5 and, citing OpenAI, 64.6% for GPT-6 Astra. Those are lab-run numbers with a standard error of 3.5 to 5 points per model, and Anthropic’s own setup reproduced Opus 5 at 29.0% against the leaderboard’s 30.0%.