AutoResearchBench - Deep Research: leaderboard

Deep Research on AutoResearchBench (2026): a ReAct agent with a scientific literature search tool must track down the single paper that satisfies a progressively specified, multi-step query, within 30 turns. Accuracy in percent; higher is better.

Metric: Accuracy (%) of pinpointing the single target paper on the Deep Research queries (progressive multi-step probing over full-text scientific literature), a ReAct agent with the DeepXiv scientific-literature search tool, at most 30 turns; higher is better. Source: arxiv.org. Saturation forecast: Around August 2028. 13 models tracked. Updated 7 Oct 2026.

Top models

#ModelSettingScoreOverall rank
1Claude Opus 4.69.39#113
2Gemini 3.1 Pro (Preview)7.93#99
3GPT-5.47.44#135
4Qwen 3.5 397B A17B6.97#264
5Claude Sonnet 4.66.96#173
6Seed 2.0 Pro6.8#222
7Kimi K2.54.69#266
8DeepSeek V3.24.21#419
9Qwen 3.5 122B A10B3.88#346
10Qwen 3 Max3.24#415
11MiniMax-M2.52.91#600
12Gemini 3 FlashNon-reasoning2.75#175
13Qwen 3.5 35B A3B1.94#492

No result here: #1 Claude Fable 5.2, #3 Gemini 4 Neon, #4 GPT-6.2 Astra.

Interactive version: theaggregate.ai/benchmark?slug=autoresearchbench-deep-research · How It Works · Data refreshed daily, snapshot 2026-11-21.