Wiki Live Challenge: leaderboard
Deep research agents write a Wikipedia-style article on the topics of 100 Wikipedia Good Articles created between March and December 2025; Wiki Live Challenge (Wang et al., 2026). A Gemini-2.5-pro judge compares each generated article with the Good Article on 39 Good Article criteria (well-written, neutral, broad in coverage), and the score is the percentage of comparisons the agent's article wins; higher is better.
Metric: Wiki Writing win rate (%; share of 39 Good Article criteria x 100 articles on which a Gemini-2.5-pro judge prefers the generated article to the Wikipedia Good Article). Source: arxiv.org. Saturation forecast: Rough model projection: around 2026. 11 models tracked.
Top models
| # | Model | Score | Overall rank |
|---|---|---|---|
| 1 | Gemini 3 Pro Deep Research | 58.33 | |
| 2 | langchain-open-deep-research-gpt-5 | 53.62 | |
| 3 | Gemini 2.5 Pro Deep Research | 35.18 | |
| 4 | OpenAI o3 Deep Research | 31.08 | |
| 5 | Grok Deep Search | 28.38 | |
| 6 | Perplexity Deep Research | 27.38 | |
| 7 | Qwen3-Max Deep Research | 25.15 | |
| 8 | langchain-open-deep-research | 20.67 | |
| 9 | Doubao Deep Research | 19.13 | |
| 10 | Tongyi-DeepResearch-30B-A3B | 15.05 |
Interactive version: theaggregate.ai/benchmark?slug=wiki-live-challenge · How It Works · Data refreshed daily, snapshot 2026-11-12.