DeepResearch Bench II: leaderboard
DeepResearch Bench II (Li et al., 2026) gives deep research agents 132 report-writing tasks in 22 domains derived from expert-written open-access review articles, and checks each report against 9,430 binary rubrics extracted from those articles and refined by domain experts. A Gemini-2.5-Pro judge marks each rubric satisfied or not; the score is the share of rubrics passed in percent over information recall, analysis and presentation, higher is better.
Metric: Rubric pass rate (%; 132 report tasks in 22 domains; share of expert-derived binary rubrics a report satisfies, Gemini-2.5-Pro judge; all three dimensions pooled). Source: arxiv.org. Saturation forecast: Rough model projection: around 2026. 8 models tracked.
Top models
| # | Model | Score | Overall rank |
|---|---|---|---|
| 1 | o3 Deep Research | 45.4 | |
| 2 | Gemini 3 Pro Deep Research | 44.6 | |
| 3 | Gemini 2.5 Pro Deep Research | 41.98 | |
| 4 | Doubao Deep Research | 40.99 | |
| 5 | Qwen3-Max Deep Research | 39.25 | |
| 6 | Grok Deep Search | 39.23 | |
| 7 | Perplexity Deep Research | 38.58 | |
| 8 | Tongyi-DeepResearch-30B-A3B | 29.89 |
Interactive version: theaggregate.ai/benchmark?slug=deepresearch-bench-ii · How It Works · Data refreshed daily, snapshot 2026-11-12.