AgentWebBench (LLM-Selected Tools) - Deep Research: leaderboard
AgentWebBench (2026) models the Agentic Web: a user agent serves information needs over 100 ClueWeb22-B websites by coordinating with content agents or tools that each cover one website, with every LLM in thinking mode. This board is key point recall in percent of long research reports on DeepResearchGym queries, under LLM-selected tools, the user agent choosing websites from their descriptions and querying a per-site retrieval tool; higher is better.
Metric: Key Point Recall (%) of long reports on 331 DeepResearchGym queries covered by the corpus, LLM user agent under tool-based prompted selection (the user agent chooses websites from their descriptions and calls a per-website dense retrieval tool, no content agents), thinking mode on; higher is better. Source: arxiv.org. Saturation forecast: Around February 2027. 7 models tracked. Updated 7 Oct 2026.
Top models
| # | Model | Setting | Score | Overall rank |
|---|---|---|---|---|
| 1 | GPT-5 Mini | 70.62 | #364 | |
| 2 | DeepSeek V3.2 | Thinking | 64.23 | #419 |
| 3 | Gemini 3 Flash (Preview) | 62.38 | #152 | |
| 4 | Qwen 3 Next 80B A3B (Thinking) | 57.85 | #594 | |
| 5 | Qwen 3 30B A3B 2507 (Thinking) | 54.82 | #764 | |
| 6 | Qwen 3 14B | 49.12 | #1269 | |
| 7 | Qwen 3 4B 2507 (Thinking) | 47.83 | #1433 |
No result here: #1 Claude Fable 5.2, #3 Gemini 4 Neon, #4 GPT-6.2 Astra.
Interactive version: theaggregate.ai/benchmark?slug=agentwebbench-llm-selected-tools-deep-research · How It Works · Data refreshed daily, snapshot 2026-11-21.