Terminal-Bench-Science Opens for Contributions: 100+ Science Tasks Frontier Agents Can Only Solve 10-20% of
Terminal-Bench is extending into the natural sciences. Terminal-Bench-Science (TB-Science) is a new benchmark targeting the end-to-end computational workflows that practicing scientists run in research labs, and the team behind it is calling for domain expert contributions to reach 100+ tasks before release.
The base Terminal-Bench has been adopted by OpenAI, Anthropic, and Google DeepMind as a measure of coding agent capability. TB-Science applies the same containerised, programmatically verified approach to scientific computing.
Why Existing Benchmarks Fall Short
Current AI-for-Science benchmarks measure textbook knowledge and abstract reasoning: hypothesis generation, literature summarisation, conceptual problem solving. None measure whether an AI agent can execute the actual computational pipelines scientists depend on daily — data analysis, simulation setup, numerical solvers, instrument data processing, model fitting, and signal processing.
TB-Science closes that gap by porting real workflows from active research labs into executable tasks. Evaluation runs in containerised environments with deterministic, programmatic verification against concrete outputs: numerical results, generated files, statistical fits, or reproducible data.
Four Domains, 100+ Tasks
The benchmark targets tasks across four science areas:
| Domain | Subfields |
|---|---|
| Life Sciences | Biology, Medicine, Neuroscience |
| Physical Sciences | Astronomy, Chemistry & Materials, Physics |
| Earth Sciences | Atmospheric Science, Geoscience, Ocean Science |
| Mathematical & Computational Sciences | Applied Mathematics, Scientific Computing, Data Science & Statistics |
The 10-20% Target
TB-Science sets an explicit difficulty floor: frontier models should be able to complete only 10-20% of tasks at the benchmark’s release. That target is deliberate — it mirrors the design philosophy of Terminal-Bench, which resets the bar whenever saturation approaches. Terminal-Bench 4.0 currently shows frontier agents at 37-70% solve rates depending on the model; TB-Science is designed to land an order of magnitude harder at launch.
Tasks must meet three criteria: drawn from real research workflows rather than textbook problems; producing verifiable, concrete outputs rather than open-ended text; and difficult enough that no current frontier agent can reliably solve them.
Status
The project launched internally in Q1 2026, opened for community contributions in Q2, and is targeting release for Q4 2026. Contributions come through a curated process via the project’s Discord. The public release will include a leaderboard and a paper submission to accompany it.
The goal is a feedback loop between practicing scientists who define what matters in scientific computing and the frontier labs building general-purpose AI agents — the same dynamic that made Terminal-Bench an effective forcing function for coding agent progress over the past 18 months.