E2E-Bench-Hard (AstaBench 2026)
Science · 2026-04-30
Unscaffolded end-to-end scientific discovery scored for partial completion using the stricter replacement judge announced in April 2026. This score is not the fraction of perfectly completed workflows. Shares a factor group with the scaffolded variant.
Top models (higher is better)
| Model | Score |
|---|---|
| Opus 4.7 | 0.7 |
| Sonnet 4.6 | 0.6 |
| Opus 4.6 | 0.6 |
| GPT-5.5 | 0.4 |
| GPT-5 | 0.4 |
| Gemini 3.1 Pro Preview | 0.3 |
| Sonnet 4 | 0.3 |
| GPT-5.4 | 0.2 |
| o3 | 0.2 |
| GPT-5 Mini | 0.1 |
| GPT-4.1 | 0.1 |
| Haiku 3.5 | 0.0 |
| GPT-4o (2024-08-06) | 0.0 |