SciDocBench: leaderboard
Workflow-centred scientific document understanding: overall score over 496 evaluation instances derived from 124 questions on arXiv, bioRxiv and OpenReview papers, spanning seven capability groups. Higher is better. Source: arxiv.org/abs/2609.05141.
Metric: Score (%). Source: arxiv.org. Saturation forecast: Around March 2028. 14 models tracked.
Top models
| # | Model | Score | Overall rank |
|---|---|---|---|
| 1 | Claude Opus 5 | 62.6 | #16 |
| 2 | GPT-5.6 Sol | 61 | #22 |
| 3 | GPT-5.6 Terra | 54.19 | #40 |
| 4 | Claude Opus 4.8 | 53.41 | #39 |
| 5 | GPT-5.6 Luna | 50.58 | #79 |
| 6 | Kimi K2.5 | 50.38 | #152 |
| 7 | Claude Sonnet 4.6 | 49.57 | #94 |
| 8 | Claude Haiku 4.5 | 40.35 | #287 |
No result here: #3 Claude Opus 5.5, #6 GPT-6 Astra, #7 GPT-6.1 Sol.
Interactive version: theaggregate.ai/benchmark?slug=scidocbench · How It Works · Data refreshed daily, snapshot 2026-11-05.