Science's Last Exam tests AI agents on research across nine fields
The benchmark tests agents' research methods and software use; every model reportedly scored zero on its end-to-end research track.
TLDR
Science's Last Exam tests whether agents can form hypotheses, plan protocols, run scientific software and interpret experiment videos. One post says researchers from more than 30 institutions wrote and reviewed the tasks. It reports that the strongest configuration, Claude Opus 5 in Claude Code, scored 37.5 out of 100 on the initial leaderboard, while every model scored zero on the end-to-end research track.
Combined views
—
2 Sources, first seen ago
