DeepResearchBench: leaderboard
Evaluates deep research agents on multi-step information gathering and synthesis tasks. Tests ability to conduct thorough research across complex topics.
Metric: Average Score. Source: deepresearch-bench.github.io. Status: years away from saturation. 41 models tracked.
Top models
| # | Model | Score |
|---|---|---|
| 1 | Claude Opus 4.6 (High) | 55.3 |
| 2 | Claude Sonnet 4.6 (High) | 54.9 |
| 3 | Claude Opus 4.5 (20251101) (High) | 54.8 |
| 4 | GPT-5.5 (High) | 54 |
| 5 | Claude Opus 4.8 (High) | 50.2 |
| 6 | GPT-5 (Low) | 49.6 |
| 7 | GPT-5.5 (Medium) | 49.6 |
| 8 | Claude Opus 4.8 (Low) | 49.3 |
| 9 | Gemini 3 Flash | 49 |
| 10 | GPT-5 (Minimal) | 48.9 |
| 11 | GPT-5.5 (Low) | 48.7 |
| 12 | GPT-5 (Medium) | 48.6 |
| 13 | Claude Opus 4.1 (20250805) | 48.3 |
| 14 | GPT-5 (High) | 48.1 |
| 15 | Gemini 3 Flash (Preview) (High) | 47.9 |
Interactive version: theaggregate.ai/benchmark?slug=deepresearchbench · How It Works · Data refreshed daily, snapshot 2026-09-05.