A new benchmark finds that today's best AI science agents finish only about one in five real research workflows end to end.
FrontierChallenge is a cross-domain benchmark of 300 planned scientific workflows spanning quantum chemistry, molecular dynamics, materials characterization, analytical chemistry, life science, and electrochemistry/environment. Researchers released and tested 97 of those tasks against twelve frontier models running under three different agent scaffolds. The best-performing setup completed just 20 of the 97 tasks, a 20.6% pass rate. Partial-progress scores told a misleading story: analytical chemistry tasks averaged an 87.6 partial score but only a 4% pass rate, and electrochemistry/environment tasks averaged 94.9 with a 0% pass rate.
That gap between looking close and actually finishing is the real story. Most agent benchmarks reward partial credit, which can make a system look nearly research-ready when it hasn't produced a single complete deliverable. The researchers also found that 75.5% of non-passing Claude Code runs still ended with the agent claiming the task was done - a mismatch between confidence and correctness that matters if a scientist is relying on the output unsupervised.
It's the same lesson AI coding benchmarks learned the hard way: an agent that sounds finished and an agent that is finished are not the same thing, and right now most of them can't tell the difference.