Terminal-Bench-Science benchmark released to evaluate AI agents on scientific workflows

The release of Terminal-Bench-Science 0.1 introduces a benchmark that measures AI agent performance on scientific workflows drawn directly from researchers’ own work. It shifts evaluation from abstract exercises to realistic tasks such as data analysis, simulation, theorem proving, and image reconstruction. The benchmark is led by Stanford researchers and built with domain experts from multiple scientific disciplines. Its first release contains 70 tasks covering life, physical, Earth, mathematical, and engineering sciences.

Adopting the benchmark requires engineers to run agents in environments that can produce concrete artifacts like code, data products, or proofs. Evaluation depends on reproducible, task-specific tests that grade those artifacts. The benchmark covers diverse workflows across life, physical, Earth, mathematical, and engineering sciences. The benchmark is designed to evolve with regular releases that allow new workflows to be added.

The benchmark stops being informative when tasks become too easy for frontier systems, as the selection process excludes workflows that today’s models already solve easily. The benchmark’s scope is limited to the five scientific domains represented in its initial 70 tasks. Evaluation depends on reproducible, task-specific tests that grade concrete artifacts such as analyses, simulations, proofs, code, and data products.

Only one feed carried the announcement, so there is no cross-source corroboration to validate the reported details. The benchmark’s usefulness will depend on the community’s ability to propose, review, and integrate new workflows over time. The benchmark is intended to evolve alongside the AI frontier through regular releases.