Science’s Last Exam finds frontier agents still fail end-to-end research
The benchmark gives AI agents real scientific tools and hidden expert checks; the strongest composite result was 37.5 out of 100, while every tested system scored zero on the end-to-end task.
Science’s Last Exam is a new benchmark designed to test whether AI agents can do scientific work rather than answer scientific questions. Its first release contains 113 tasks across nine disciplines, covering hypothesis discovery, protocol planning, scientific computing, software operation, experiment observation and one end-to-end research assessment.
Agents receive real research materials and tools, including specialist software, recorded lab work and containerized environments. Their outputs are checked against withheld references or independent graders that remain outside the agent’s environment. The benchmark team says the strongest composite score among eight evaluated frontier systems was 37.5 out of 100. Every tested system scored zero on the current end-to-end task.
That result is useful, but it needs context. The release contains only one end-to-end task, that section is explicitly marked for expansion, and wet-lab execution is still under construction. The benchmark therefore does not prove that AI cannot perform any complete scientific workflow; it shows that current agents remain brittle on a demanding, early test.
The evaluation framework is available on GitHub, with task code, graders and recorded runs. Public data downloads are opening separately, while some hidden reference assets require approved access. The project gives researchers a more practical way to compare scientific agents on verifiable work instead of polished final answers.
