CORE-Bench OOD (Codex CLI): leaderboard
CORE-Bench OOD (2026) applies the CORE-Bench v1.1 reproduction task to 19 capsules from physics, engineering, economics and computer science, fields outside the original benchmark. This board is task accuracy in percent with the Codex CLI scaffold; higher is better.
Metric: Task accuracy (%) on the 19 CORE-Bench OOD tasks (physics, engineering, economics and computer science capsules outside CORE-Bench's fields): reproduce a paper's computational results from its code capsule and answer the task questions; Codex CLI scaffold with default settings (subagents off), reasoning effort as labeled; higher is better. Source: arxiv.org. Saturation forecast: Estimated already saturated. 8 models tracked.
Top models
| # | Model | Score | Overall rank |
|---|---|---|---|
| 1 | GPT-5.2 (Medium) | 100 | #176 (GPT-5.2) |
| 2 | GPT-5.4 (xHigh) | 100 | #123 (GPT-5.4) |
| 3 | GPT-5.1 (Medium) | 94.7 | #229 (GPT-5.1) |
| 4 | GPT-5.4 (High) | 89.5 | #123 (GPT-5.4) |
| 5 | GPT-5 (Medium) | 89.5 | #167 (GPT-5) |
| 6 | GPT-5.4 (Medium) | 89.5 | #123 (GPT-5.4) |
| 7 | GPT-5.3 Codex (Medium) | 89.5 | #116 (GPT-5.3 Codex) |
| 8 | GPT-5.4 (Low) | 84.2 | #123 (GPT-5.4) |
No result here: #1 Claude Fable 5.2, #2 Gemini 4 Carbon, #5 Claude Opus 5.5.
Interactive version: theaggregate.ai/benchmark?slug=core-bench-ood-codex-cli · How It Works · Data refreshed daily, snapshot 2026-11-13.