CORE-Bench v1.1 (CORE-Agent): leaderboard
CORE-Bench v1.1 (2026) is CORE-Bench Hard with 15 task errors and 20 shortcut tasks fixed and 10 new tasks: 39 tasks in which an agent reproduces the computational results of a paper from its code capsule and answers questions about them. This board is task accuracy in percent with the CORE-Agent scaffold; higher is better.
Metric: Task accuracy (%) on the 39 CORE-Bench v1.1 tasks (CORE-Bench Hard with 15 task errors and 20 shortcut tasks fixed plus 10 new tasks): reproduce a paper's computational results from its code capsule given only the README, code and data, and answer the task questions; the CORE-Agent scaffold of the original benchmark, reasoning as labeled; higher is better. Source: arxiv.org. Saturation forecast: Rough model projection: around 2026. 3 models tracked.
Top models
| # | Model | Score | Overall rank |
|---|---|---|---|
| 1 | CORE-Agent + Claude Opus 4.6 (Non-reasoning) | 100 | |
| 2 | CORE-Agent + Claude Opus 4.5 (Non-reasoning) | 82.1 | |
| 3 | CORE-Agent + GPT-5.4 (Medium) | 51.3 |
Interactive version: theaggregate.ai/benchmark?slug=core-bench-v1-1-core-agent · How It Works · Data refreshed daily, snapshot 2026-11-13.