HAL CORE-Bench Hard — leaderboard
Princeton HAL leaderboard for CORE-Bench Hard: 270 computational reproducibility tasks from 90 scientific papers in CS, social science, and medicine. Tests ability to reproduce research results.
Metric: Accuracy (%). Source: hal.cs.princeton.edu. Status: saturation imminent. 49 models tracked.
Top models
| # | Model | Score |
|---|---|---|
| 1 | Claude Opus 4.5 | 77.78 |
| 2 | Claude Sonnet 4.5 | 62.22 |
| 3 | Claude Opus 4.1 (20250805) | 51.11 |
| 4 | Claude Sonnet 4 (20250514) | 46.67 |
| 5 | Claude Sonnet 4.5 (High) | 44.44 |
| 6 | Claude Opus 4.1 | 42.22 |
| 7 | Claude Opus 4.5 (20251101) | 42.22 |
| 8 | Claude Opus 4.5 (High) | 42.22 |
| 9 | Claude Opus 4.1 (High) | 42.22 |
| 10 | Gemini 3 Pro (Preview) (High) | 40 |
| 11 | Claude 3.7 Sonnet (High) | 37.78 |
| 12 | Claude 3.7 Sonnet (20250219) | 35.56 |
| 13 | O4 Mini (High) | 35.56 |
| 14 | GPT-4.1 (2025-04-14) | 33.33 |
| 15 | GPT-5 (Medium) | 26.67 |
Interactive version: theaggregate.ai/benchmark?slug=hal-core-bench-hard · How the rankings work · Data refreshed daily, snapshot 2026-07-22.