AI Agents in Science: The Best Solves Just 30% of Tasks

Stanford HAI, together with the Allen Institute, NSF IFML, Ai2, and 376 researchers from 22 countries, has published Terminal-Bench-Science — the first benchmark where tasks come from scientists' actual working projects rather than being purpose-built for testing. The 70 tasks span five domains: life sciences (biology, medicine, neuroscience), physical sciences (physics, chemistry, materials science), Earth sciences, mathematics, and engineering.

1Results: 30% Is the Ceiling

Claude Opus 5 led with a score of 30.0%. GPT-5.6 Sol reached 22.4%, Claude Fable 5 scored 21.4%, and Claude Opus 4.8 scored 10.5%. All other models, including Grok 4.6 and Kimi K3, failed to exceed 8%. The breakdown by discipline is uneven: in mathematics, Claude Fable 5 (33.3%) and GPT-5.6 Sol (31.4%) lead the pack, while in physical sciences and life sciences Claude Opus 5 holds first place with scores around 28–30%.

2Why These Tasks Are Harder Than Standard Tests

The authors initially received 920 submissions, of which only 70 — fewer than 8% — passed screening. Each task had to be scientifically meaningful, sufficiently challenging for frontier models, and unambiguously verifiable. Task types include data analysis, simulations, optimization, theorem proving, image reconstruction, and machine learning. These are not multiple-choice questions — the agent works in a terminal and must produce a concrete result.

3Cost: More Expensive Than It Looks

Running the benchmark is not cheap. Claude Opus 5 consumed roughly 12 billion tokens and $7,000 to solve 30% of tasks. GPT-5.6 Sol proved more cost-efficient: $4,200 for 22.4%. Claude Fable 5 was the most expensive — $14,200 for 21.4%. This is important context: today's scientific agents are not only slow at solving problems, they also consume significant resources.

The benchmark is designed as a living standard: the next version, 0.2, is already in development, with a pull request deadline of October 5, 2026. The authors emphasize that the bar will rise alongside model capabilities. The goal is not to document AI's shortcomings but to give the industry an honest mirror for measuring real scientific progress.