EurekaBench Predictive Accuracy
Science · 2026-09-30
Predictive accuracy of the mechanisms AI agents submit on EurekaBench's 26 problems: each submitted implementation is run on held-out test data and every prediction criterion receives a normalized continuous score from 0 to 1. Reported without the scientific-constraint gate and averaged over problems; a component of the conditioned score.
Top models (higher is better)
| Model | Score |
|---|---|
| Claude Fable 5.1 | 47.4 |
| GPT-6 Astra | 47.4 |
| Claude Opus 5 | 42.9 |
| Opus 4.8 | 42.7 |
| GPT-5.6 Sol | 41.9 |
| Kimi K3 | 28.8 |