EurekaBench Scientific Insights
Science · 2026-09-30
Share of EurekaBench's 306 expert-written insight questions, over 26 problems, whose insight an agentic judge can derive from the agent's submitted mechanism with supporting evidence, optionally by running further simulator experiments without changing the mechanism. Reported without the scientific-constraint gate, averaged over two judges and over problems; a component of the conditioned score.
Top models (higher is better)
| Model | Score |
|---|---|
| Claude Fable 5.1 | 42.4 |
| Claude Opus 5 | 36.2 |
| GPT-6 Astra | 29.4 |
| Opus 4.8 | 26.9 |
| GPT-5.6 Sol | 25.9 |
| Kimi K3 | 25.2 |