Stanford-led Terminal-Bench-Science tests AI agents on 70 real scientific workflows

A group led by Stanford researchers, working with the team behind Terminal-Bench and domain experts from research institutions around the world, has announced Terminal-Bench-Science, a benchmark that scores AI agents on workflows taken from practicing scientists' own work. Version 0.1 ships with 70 tasks spanning the life, physical, Earth, mathematical and engineering sciences, and the strongest system evaluated resolves only 30% of them.

What the agents are asked to do

Rather than quiz questions or standardized exercises, the tasks reproduce genuine research work: analyzing scientific data, statistical inference, simulation, optimization, theorem proving, image reconstruction, signal processing, inverse problems, sensor calibration, model fitting, classification and scientific machine learning. Agents run in realistic environments and are graded on concrete outputs — code, analyses, proofs, simulations and data products — using reproducible tests written for each task.

The initial pool leans toward the life sciences with 19 tasks, followed by 17 each in the physical and mathematical sciences, nine in engineering and eight in Earth sciences.

The opening leaderboard

According to the benchmark's announcement, every system was run three independent times per task across the full set. Claude Opus 5, paired with the Claude Code harness, leads at a 30.0% resolution rate. GPT-5.6 Sol with Codex follows at 22.4%, with Claude Fable 5 close behind at 21.4%. Claude Opus 4.8 manages 10.5%, while GPT-5.6 Terra (8.6%), GLM 5.3 (8.1%), Kimi K3 (7.1%), Grok 4.6 (7.1%) and GPT-5.6 Luna (3.3%) fill out the lower ranks. GLM 5.3 is the strongest open model in the field.

Cost and efficiency add nuance. Reaching the top score cost roughly $7.0k in total evaluation spend for Opus 5, while GPT-5.6 Sol delivers the same resolution as Claude Fable 5 for under a third of the price ($4.2k versus $14.2k). Measured in tokens, Fable 5 matches Sol's performance while consuming about a quarter less (6.4 billion versus 8.4 billion). Only Kimi K3 and Claude Opus 5 land on the efficiency frontier of both the cost and token plots.

A funnel built to be hard to pass

Tasks arrive through an open process on GitHub, with discussion in a dedicated Discord channel. Ideas begin as proposals that reviewers debate; those deemed a good fit are implemented as pull requests, which must then be shown to be objectively verifiable, genuinely difficult, and not already solvable by current frontier systems. Domain reviewers judge scientific validity and realism, technical reviewers inspect how each task is constructed and verified, and a final quality gate closes the process. The numbers show how selective this is: of 920 proposals, 464 were approved for implementation and 386 pull requests were opened, yet only 70 tasks made it into release 0.1. The team attributes that attrition to how hard it is to write tasks that are scientifically meaningful, tough for frontier agents and specified precisely enough to grade rigorously.

Deliberately harder than its software sibling

Terminal-Bench has tracked progress in coding agents; the science edition separates systems about as effectively while cutting resolution rates by more than 10 percentage points for every model evaluated on both — Claude Fable 5 drops from 33.8% to 21.4%, for example. The announcement describes that gap as intentional, with tasks calibrated during review to challenge the newest frontier models.

The project is also designed to avoid the fate of benchmarks released once and abandoned: it is meant to run continuously, with regular releases letting scientists contribute new workflows and refine existing ones as models advance, creating a feedback loop between research needs and AI development.

Why it matters

Most AI capability measurements are effectively set by the organizations building the models. Here the yardstick comes from the scientific community, which decides which workflows matter and what counts as success. That shift matters for credibility: agents are judged on work researchers actually do rather than on proxies for it, and the evaluation reflects the priorities of scientists rather than outside interests.

The results also calibrate expectations. A 30% ceiling on carefully vetted tasks means today's agents remain far from autonomously running demanding scientific workflows. But the project's stated aim is not full autonomy — it is agents that take over time-consuming technical work so researchers can concentrate on framing questions, forming hypotheses and validating results. A durable, evolving benchmark gives both labs and scientists a shared, honest measure of how close that goal is.