Stanford's New AI Science Test: Best AI Only Nails 30% of Tasks

Scientists at Stanford University have built a new way to test whether AI can genuinely help with scientific research. Instead of asking AI to answer textbook questions, they asked working scientists to submit real tasks from their own labs — things like analyzing biological data, running climate simulations, solving math proofs, and calibrating sensors. The result is Terminal-Bench-Science, a living test that keeps evolving as AI improves.

So far, the results show we're still early. The best AI tested, Claude Opus 5, successfully completed only 30% of the 70 tasks. That means most real scientific work is still beyond AI's reach. But the goal isn't to replace human scientists — it's to give them capable research assistants that handle the tedious, time-consuming grunt work.

Why should you care? Because science drives the medicines, clean energy, and climate solutions you rely on. If AI can take over the slow parts of research — like running a thousand simulations or sorting through massive datasets — scientists can spend more time on big-picture thinking, like coming up with new hypotheses and deciding what experiments matter most. That could speed up breakthroughs.

The clever part is who controls the test. Working scientists from many fields contributed the tasks, not tech companies or data vendors. That means AI progress is measured by what researchers actually need, not by what AI companies want to show off. As new AI models come out, scientists will keep adding harder tasks, creating a loop where AI must chase real scientific usefulness rather than test scores.