KDR-Bench: leaderboard
Report quality on KDR-Bench (2026), a knowledgeable deep research benchmark of 41 expert-level questions in nine domains whose answers need web content and structured statistics tables. Reports are scored with the RACE protocol, the equal-weight mean of comprehensiveness, depth, coherence and readability judged against a reference report, from 0 to 100 with 50 meaning parity; higher is better.
Metric: RACE general-purpose score (0-100): equal-weight mean of comprehensiveness, depth, coherence and readability, each scored against a reference report (50 means parity with the reference), on the 41 expert-level research questions of KDR-Bench in nine domains backed by 1,252 Statista tables, each report judged by DeepSeek-V3.2; higher is better. Source: arxiv.org. Saturation forecast: Around February 2027. 9 models tracked. Updated 7 Oct 2026.
Top models
| # | Model | Score | Overall rank |
|---|---|---|---|
| 1 | Gemini 3 Pro Deep Research | 50.2 | |
| 2 | Perplexity Deep Research | 46.5 | |
| 3 | MiniMax-M2 | 44.7 | #689 |
| 4 | Qwen 3 Max | 44.6 | #415 |
| 5 | Grok 4 DeeperSearch | 44.5 | |
| 6 | Tongyi DeepResearch 30B A3B | 41.8 | |
| 7 | Deep Research | 41.3 | |
| 8 | GLM-4.6 | 40.8 | #498 |
| 9 | Hunyuan 2.0 | 40.6 | #690 |
No result here: #1 Claude Fable 5.2, #3 Gemini 4 Neon, #4 GPT-6.2 Astra.
Interactive version: theaggregate.ai/benchmark?slug=kdr-bench · How It Works · Data refreshed daily, snapshot 2026-11-21.