AItonomy Foundation launches Science's Last Exam, a 113-task AI research benchmark

AIAItonomy Foundation introduces Science's Last Exam, a benchmark that gives agents research goals and real-world tools, then grades them against hidden expert checks. Of eight frontier models tested, the best scored 37.5 out of 100, and all scored 0 on true end-to-end research. The benchmark covers 113 tasks across 9 disciplines and opens on Nov 6.

AIScience's Last Exam is a new benchmark built from 113 real research tasks across nine scientific fields, written and reviewed by researchers at more than 30 institutions. On the initial leaderboard, the strongest configuration, Claude Opus 5 in Claude Code, scores 37.5 out of 100, while every model scores 0 on the end-to-end research track.