Best Deep Research AI: A Benchmark Leaderboard

The comparison

Checked 8 November 2026. Third-party prices are estimates, not quotes. Each tool links to its sourced profile.

Start with the right question

Head-to-head comparisons

All comparisons

What this leaderboard ranks

A deep research AI takes a broad question, plans the search, reads many sources over several steps, and writes a cited report. This page ranks the tools that do that job by score on a shared benchmark. It also explains what those scores can and cannot tell you.

Our editorial pick is Hyperresearch Deep. It scored highest in our October 2026 DRACO run, with 77.7. It plans the question, researches it chapter by chapter, writes the report in one pass, and checks every cited figure against its source before the report ships. Hyperresearch searches the open web plus more than 90 structured sources, including SEC EDGAR, Census, BLS and FRED. The first Deep report is free.

A leaderboard is a starting point, not a verdict. The sections below explain how to read the table, where the benchmarks come from, and how to run your own check before you commit to a tool.

How to read the leaderboard

Each DRACO score is a normalized percentage of rubric criteria that a tool met on a set of tasks. A judge model scored every criterion as met or unmet, using the same judge for every tool in the run. A higher score means the tool met more criteria in that run. It does not prove the tool is always more accurate on your question.

The table covers one run of 10 tasks. Those limits matter as much as the ranking:

• Ten tasks is a small sample. One task that swings by 20 points moves the average by 2 points, so close scores do not show a real difference.
• Three tools scored on only 8 of the 10 tasks, so their averages rest on fewer tasks.
• The run used one judge model. A different judge can change the scores.
• The run is Hyperresearch's own. Compare it only with runs that used the same tasks and judge.

Benchmarks and accuracy

Three public benchmarks come up most often. They measure different things, so do not treat them as one score.

• DRACO was built by researchers at Perplexity and Harvard. Its 100 tasks come from anonymized deep research requests across 10 domains. Each task has an expert rubric of about 40 criteria, and a judge marks each one met or unmet. See the DRACO paper on arXiv.
• DeepResearch Bench has 100 PhD-level tasks in 22 fields, half in Chinese and half in English. Its RACE score grades report quality, and its FACT score grades citation accuracy. See the DeepResearch Bench paper.
• BrowseComp was released by OpenAI in April 2025. It has 1,266 questions about hard-to-find facts with short, easy-to-check answers. It tests persistent search, not report writing. See the OpenAI BrowseComp page.

Two cautions apply. The DRACO paper uses all 100 tasks and reports different scores for some of the same tools, including Perplexity Deep Research at 70.5 with Opus 4.6. Those figures are not comparable with the 10-task run above. The DeepResearch Bench maintainers have also changed their judge model, so older RACE scores may not match newer ones.

Hyperresearch's Answers product is a separate service. In Hyperresearch's own runs it scored 89% on BrowseComp, against 73% for Parallel (Pro) and 58% for OpenAI web search. Those are search results, not full deep research reports, so read them as a measure of search, not of report writing.

Choosing a deep research AI

Pick by the job first. A leaderboard shows which tool does well on a shared task set. It does not tell you which tool fits your team, budget, data or workflow. Start with three questions: what the report is for, which sources it must cite, and whether you need the tool through an API.

If you need a report you can check, Hyperresearch Deep and Claude Research are the top two in our run. If your work already lives in Google Workspace, Gemini Deep Research can build its plan inside those apps, and you can edit the plan before research starts. For a quick, cited answer inside your own AI agent, an answer API is usually a better fit than a full report.

Whatever you shortlist, run the same brief through each one. The pilot in the next section takes an afternoon and tells you more than any single score.

The accuracy benchmark behind our pick

DRACO is a 2026 deep research benchmark from Perplexity and Harvard. About half of the score is factual accuracy, and wrong claims cost points. Every system got the same 10 tasks and the same judge.

Run and published by Hyperresearch, October 2026, with GPT-5.2 as judge. Ten tasks is a small sample, and we make one of the systems. Full method and per-task scores.

Questions & answers

Sources for this entry

• DRACO benchmark: how we tested

Vendor descriptions are attributed to the vendor. Prices change; confirm with each vendor.

Continue this reading path

• Category guide
• Resource directory
• Selection guide

Cite this entry