Terminal-Bench-Science leaderboard: only two models clear 50%

Terminal-Bench-Science 0.1 tests whether an AI agent can carry out a piece of scientific research work from start to finish, and the best models currently pass a little under two-thirds of its tasks. Artificial Analysis launched a leaderboard for the benchmark, which was built by Steven Dillmann and researchers at Stanford with the Terminal-Bench team and open source scientific contributors. It has 70 expert-curated tasks based on real research in five domains: life sciences, physical sciences, mathematical sciences, engineering sciences and earth sciences. Each task drops the agent into a sandbox with the data, tools and instructions it needs, and automated tests grade the result pass or fail. The reported score is the share of tasks passed on a single attempt, averaged over three runs.

Only two models score above 50%: GPT-6 Astra at 63% on its max setting, and Claude Opus 5.5 at 62% on xhigh and 59% on max. The best open-weights models score far lower, with GLM-5.3 at 10% and DeepSeek V4.1 Flash at 9%, more than 50 points behind the leaders. Artificial Analysis says recent releases have made large advances but that even GPT-6 Astra has significant headroom.

The breakdown by field is suggestive but rough. Life sciences is the lowest-scoring domain for most leading models, and Claude Opus 5.5 on xhigh passes 71% of mathematical sciences tasks but 46% of life sciences tasks. Each domain has only 8 to 19 tasks, so a task or two can move a domain score noticeably, and Artificial Analysis itself calls these scores noisy. Stanford AI Lab's release post for the benchmark is the companion item on the benchmark itself.