DeepResearchEval - Factual Accuracy: leaderboard
DeepResearchEval (Wang et al., 2026) fact-checks the reports that deep-research systems write for 100 research tasks: a GPT-5-mini agent with web search extracts checkable statements (numbers, events, dates, people) and verifies each against retrieved evidence. Score is the percentage of checkable statements judged correct, with wrong and unverifiable statements counting against; higher is better.
Metric: Verified-correct share (%; checkable statements a GPT-5-mini web-search agent judges right, of all checkable statements; wrong or unverifiable count against). Source: arxiv.org. Saturation forecast: Rough model projection: around 2026. 9 models tracked.
Top models
| # | Model | Score | Overall rank |
|---|---|---|---|
| 1 | Manus | 82.3 | |
| 2 | Gemini-2.5-Pro Deep Research | 76.62 | |
| 3 | DeepSeek Deep Research | 76.44 | |
| 4 | OpenAI Deep Research | 76.21 | |
| 5 | Qwen-3-235B-A22B Deep Research | 72.39 | |
| 6 | Doubao Deep Research | 69.5 | |
| 7 | Grok4 Deep Research | 61.81 | |
| 8 | DeepResearchEval Claude-Sonnet-4.5 Deep Research (checkpoint unspecified) | 60.72 | |
| 9 | Perplexity Deep Research | 58.94 |
Interactive version: theaggregate.ai/benchmark?slug=deepresearcheval-factual-accuracy · How It Works · Data refreshed daily, snapshot 2026-11-12.