PaperBench: leaderboard
End-to-end replication of AI papers, graded against hierarchical rubrics.
Metric: Replication Score (%). Source: openai.com. Source: github.com. Saturation forecast: Estimated already saturated. 6 models tracked.
Top models
| # | Model | Score | Overall rank |
|---|---|---|---|
| 1 | O1 (High) | 26 | #183 (O1) |
| 2 | Claude 3.5 Sonnet | 21 | #350 |
| 3 | O3 Mini (High) | 8.5 | #272 (O3 Mini) |
| 4 | DeepSeek R1 | 6 | #251 |
| 5 | GPT-4o | 4.1 | #343 |
| 6 | Gemini 2.0 Flash | 3.2 | #342 |
No result here: #3 Claude Opus 5.5, #6 GPT-6 Astra, #7 GPT-6.1 Sol.
Interactive version: theaggregate.ai/benchmark?slug=paperbench · How It Works · Data refreshed daily, snapshot 2026-10-20.