Observed arrival · 2026-10-24
Science’s Last Exam tests AI agents on scientific work
A research-preview benchmark that evaluates AI agents across five scientific capabilities, from experiment planning and scientific coding to hypothesis discovery.
○Open
⊠Login
$Paid
†Ads
✦Pretty
●Pro
◎Niche
◉Human
⚑Risk
ƒJS
- For
- Scientific-agent researchers and benchmark builders
- Worth noticing
- The page says Inspire Rate tracks actionable connections to later literature, not whether a generated hypothesis is a verified discovery.
Field notes
The described evaluation loop starts with scientific inputs, a target outcome, and defined tools; an agent's answer or artifact is frozen before independent checks. The page lists 154 draft Representative tasks and 44 draft Last-Exam tasks, while stating that overall scores remain withheld until coverage is complete. Hypothesis Discovery is assessed separately on research databases, and its budgets differ across systems.
Observed signals
Read the marks
Editorial observations of this landing page, not a rating.
○OpenPublic substance visible
✦PrettyNotable craft visible
◎NicheUnusually specific use
One card from the complete issue