Benchmarks
326 sources · 1,256 models × 2,956 benchmarks · 154,484 scoresScience's Last Exam
Science's Last Exam (led by AItonomy) asks whether an AI agent can carry scientific research from a question to a finding, with 113 tasks built from real scientific material in nine disciplines. Agents work in Claude Code or Codex with web search and fetch turned off, and Hypothesis Discovery runs in the DiscoveryRSI research harness. The composite is the equal-weight mean of Hypothesis Discovery, Protocol Planning, Scientific Computing, Scientific Software Operation, Experiment Observation and End-to-end research. End-to-end research is a one-task preview that scores 0 for every listed model, which caps today's composites at about 83 of 100. Wet-lab Execution isn't scored yet. Claude Opus 5.5 and GPT-6 Astra aren't evaluated yet because provider safeguards can restrict biology workflows.
No signal or difficulty estimate. Too few models have been measured on this benchmark to estimate its difficulty.
| # | Model | Setting | Score | Source |
|---|---|---|---|---|
| 1 | Claude Opus 5 | high · Science's Last Exam; Claude Code with web search and fetch off; DiscoveryRSI for Hypothesis Discovery | 37.5 | |
| 2 | GPT-6.1 Sol | medium · Science's Last Exam; Codex with web search and fetch off; DiscoveryRSI for Hypothesis Discovery | 36.1 | |
| 3 | Kimi K3.1 Pro | not reported · Science's Last Exam; Claude Code with web search and fetch off; DiscoveryRSI for Hypothesis Discovery | 28.9 | |
| 4 | MiMo-V2.6-Pro | not reported · Science's Last Exam; Claude Code with web search and fetch off; DiscoveryRSI for Hypothesis Discovery | 26.4 | |
| 5 | GLM-5.3 Flash | high · Science's Last Exam; Claude Code with web search and fetch off; DiscoveryRSI for Hypothesis Discovery | 21.3 | |
| 6 | MiniMax M3 | high · Science's Last Exam; Claude Code with web search and fetch off; DiscoveryRSI for Hypothesis Discovery | 16.8 |
