AItonomy launches Science’s Last Exam to test end-to-end research agents
113 tasks across nine disciplines score hypothesis work through observation; early runs put the best model at 37.5/100 with zero end-to-end passes.
Why it matters
Most science benchmarks ask what models know. Science’s Last Exam asks whether agents can do research: form hypotheses, plan protocols, run scientific code and software, and read instrument-like observations against hidden expert checks. The AItonomy Foundation’s public set has 113 tasks across nine fields, with a wet-lab track still under construction. Early leaderboard figures put the best of eight frontier models at 37.5 out of 100 overall and zero on true end-to-end research. Full task packages are scheduled to open on 6 November. Watch whether labs treat SLE as a training target and whether scores move without leaking the hidden checks.
Key facts
- 113 public tasks across nine disciplines
- Six capability tracks plus end-to-end assessment
- Best early score 37.5/100; zero end-to-end passes
- Led by AItonomy Foundation; site sciences-last-exam.org
- Announcement thread 2026-11-05 15:58Z; packages due Nov 6