CORE-Bench-Hard (AstaBench 2026)
Code · 2026-04-30
Computational reproducibility on the CORE-Bench-Hard AstaBench test split. This is the suite's selected test cohort, not an assumed full upstream dataset.
Top models (higher is better)
| Model | Score |
|---|---|
| GPT-5.4 | 0.9 |
| GPT-5.5 | 0.9 |
| Opus 4.6 | 0.9 |
| Gemini 3.1 Pro Preview | 0.9 |
| Opus 4.7 | 0.9 |
| Sonnet 4.6 | 0.7 |
| o3 | 0.6 |
| GPT-5 | 0.5 |
| GPT-5 Mini | 0.5 |
| Sonnet 4 | 0.4 |
| GPT-4.1 | 0.2 |
| GPT-4o (2024-08-06) | 0.1 |
| Haiku 3.5 | 0.0 |