Claude Deep Research: How Claude's Research Mode Works

Yes, Claude has deep research. Anthropic just calls it Research. Turn it on and Claude runs a chain of searches that build on each other, reads across the web and your connected apps, and comes back with a cited report instead of a quick answer.

Here's the short version:

• Where it lives: the + button in the chat box, then Research. Web search has to be on first.
• Who gets it: paid plans only. Anthropic's help center lists Pro, Max, Team and Enterprise.
• What it reads: the web, plus Google Workspace (Gmail, Calendar, Docs) and any other connectors you've enabled for that chat.
• How it differs from web search: web search is one or two lookups. Research is five or more, planned as it goes.
• Claude Code: there is an official /deep-research command. It's a built-in workflow, separate from the chat feature.

Every fact below comes from Anthropic's own help center, blog or docs. Features like this change often, so the linked pages are what to trust on the day you check.

What Claude's Research Mode Actually Does

A normal Claude chat with web search runs a lookup when it needs one, then answers. Research works differently. Anthropic's Research help article describes it as Claude running several searches that build on each other, deciding what to look into next based on what it finds, and working through open questions.

When Anthropic launched the feature, its announcement post put it this way: "Claude operates agentically, conducting multiple searches that build on each other." That's the key idea. You're not getting ten searches fired at once. You're getting a loop: search, read, notice a gap, search again.

The output is a long report with citations. For Google Workspace sources, the launch post says Claude adds inline citations so you can check where each point came from.

Which model runs Research

You don't pick the model for Research. Per the help article, it runs on Claude Opus 5.5, falling back to Opus 4.8 when needed, "regardless of the model selected for the chat." On Team and Enterprise plans, if an admin has turned off both models for your role, Research won't work at all.

How long a Research run takes

Anthropic gives two different figures, and they describe different kinds of runs:

So plan for anything from a couple of minutes to a long coffee break. Bigger questions with more sources take longer.

How to Turn On Claude Research

The setup takes about ten seconds:

1. Make sure web search is on. Research won't start without it.
2. Click the + at the bottom left of the chat box.
3. Choose Research. A blue indicator shows it's active.
4. Type your question and send it.

Click the indicator again to turn Research off.

It works in Claude on the web, Claude Desktop and the mobile apps. There's one catch on mobile, covered in the connectors section below.

If you start a Research chat and Claude seems to answer without researching, the help article's fix is simple: ask it directly to use the research tool.

Which plans include Research

Research is a paid-plan feature. The help article lists Pro, Max, Team and Enterprise. It isn't on the Free plan.

That wasn't always the case. At launch in April 2025, Research was an early beta for Max, Team and Enterprise in the US, Japan and Brazil. A June 3, 2025 update to the Integrations post opened it to Pro as well.

One more thing to know: Research uses the same usage limits as normal chats, but it burns through them faster, because it pulls many sources and writes long answers. If you're near a limit, a Research run may be what tips you over.

Claude Research vs Web Search

These two get mixed up constantly. Web search is a tool Claude reaches for during a normal answer. Research is a mode that plans a whole investigation.

Anthropic's own rule of thumb: use web search for things like checking the weather or getting recent news. Use Research for "comprehensive information gathering," such as comparing competitors, updating an outdated document, or sorting priorities from your calendar and email.

You can also combine Research with extended thinking. Anthropic suggests this for heavier jobs, like researching an emerging technology for a business proposal, where careful planning matters as much as coverage.

For the full story on how plain search behaves in the app, the API and Claude Code, see our guide to Claude web search.

Using Google Workspace and Connectors as Research Sources

This is where Claude Research stands apart from a pure web tool. It doesn't only read the internet. It can read your stuff.

Per the help article, Research searches the web and your connected tools, including Gmail, Google Calendar and Google Docs, plus any other connectors you've enabled for the chat. It uses those connector tools without asking for approval each time, which is worth knowing before you connect anything that can make changes in another app.

A few limits to plan around:

• Mobile skips connectors. Research started in the mobile app searches only the web, and it doesn't warn you. Start on the web or in Desktop if you need your files.
• No local MCP servers. Research can't use tools from local MCP servers set up in Claude Desktop.
• Team and Enterprise need admin setup. The launch post says admins must enable Google Workspace for the whole domain before individual users can connect.

If Research isn't pulling from a source you expected, name it. Something like "check my Google Docs for last quarter's pricing memo" points Claude at the right place.

Mixing private and public sources is also the main reason to pick Claude over a web-first tool. We compared the two approaches in Claude vs Perplexity.

How to Write a Good Claude Research Prompt

Research is only as good as the question. A vague prompt gets a vague report that took ten minutes to write. The tips below are practical habits, not Anthropic rules, but they follow from how the feature works: Claude plans its own searches, so the more it knows about what you want, the better its plan.

Say what decision the research is for. "Research CRM tools" is a topic. "I run a five-person agency and need to pick a CRM this month. Compare three options on price model, Gmail integration and setup time" is a brief.

Set the scope. Give a time range ("anything since January 2026"), a region, or the kinds of sources you trust ("official docs and company pages, not review sites").

Name your internal sources. If the answer depends on your own files, say which ones: "use my Q3 planning doc and recent emails from the vendor."

Ask for the format you need. A comparison table, a one-page memo, a list of open questions. Claude writes long by default in Research, so asking for structure saves you editing time.

Ask it to flag uncertainty. Add "mark anything you couldn't confirm from a primary source." That makes the weak spots easy to find.

A prompt that does all of this:

I'm choosing an email marketing tool for a small online store.
Compare three options on pricing model, Shopify integration and
deliverability features. Use only the vendors' own sites and
help docs. Put the result in a table, then list anything you
couldn't confirm from an official page.

How to Check Claude Research Citations

A cited report is not the same as a correct report. Citations make checking possible. They don't do the checking for you.

Here's a quick routine before you rely on a Research report:

1. Open the citations behind your key claims. Not all of them. Just the three or four points your decision rests on.
2. Check the claim is actually on the page. A real source can still be summarized wrong. Look for the sentence, not just the topic.
3. Check the date. Pricing, plan features and product names go stale fast. A correct 2024 page can be wrong today.
4. Watch for secondhand sources. If a stat traces back to a blog quoting another blog, find the original.
5. Ask Claude to defend weak points. "Which claims in this report rest on a single source?" is a useful follow-up.

This matters most when the report mixes your own files with the web, because a wrong number can look just as confident as a right one.

Does Claude Code Have Deep Research?

Yes, and it's official. Claude Code's commands reference lists /deep-research as a bundled workflow.

According to the workflows docs, /deep-research "fans out web searches on a question across several angles, fetches and cross-checks the sources it finds, votes on each claim, and returns a cited report with claims that didn't survive cross-checking filtered out." It needs the WebSearch tool to be available, and it only runs when you invoke it.

You run it like this:

/deep-research What changed in the Node.js permission model between v20 and v22?

A few things set it apart from Research in the chat app:

• It runs in the background. Your session stays free while subagents work. You check progress with /workflows.
• It cross-checks claims. When verifier agents can't check a claim, for example after a rate limit, the report lists it as unverified instead of counting it as wrong.
• It costs more. A workflow spawns many agents, so one run can use far more tokens than a normal turn. Runs count toward your plan's limits. Our guide to Claude Code usage limits explains how those limits work.

Be careful with search results here. You'll also find community-made "deep research" skills and repos for Claude Code. Some are good, but they're third-party. The bundled /deep-research workflow is the one Anthropic documents.

If you're still choosing a research stack beyond Claude, our roundup of AI tools for research covers the wider field.

FAQ

Does Claude have deep research?

Yes. Anthropic calls the feature Research. It runs multiple searches that build on each other and returns a cited report. In Claude Code, there's also a separate /deep-research command.

Is Claude Research free?

No. Anthropic's help center lists Research for Pro, Max, Team and Enterprise plans only. Check Claude's pricing page for what each plan includes today.

How do I turn on Claude research mode?

Turn on web search, click the + at the bottom left of the chat box, and select Research. A blue indicator appears while it's on.

Can Claude Research read my Gmail and Google Docs?

Yes, once you connect Google Workspace. Research can search Gmail, Google Calendar and Google Docs alongside the web. It won't use connectors if you start the run from the mobile app.

Why didn't Claude research my question?

Check that both web search and Research are on. If they are, ask Claude directly to use the research tool. On Team or Enterprise plans, an admin may have turned off the models Research depends on.

Is Claude Research better than Perplexity Deep Research?

It depends on your sources. Claude's edge is reading your own connected files alongside the web and writing a polished report. Web-first tools focus on broad open-web coverage.

The Short Version

• Claude deep research = Research mode. It's in the + menu, needs web search on, and is for paid plans.
• It plans its own searches, reads the web plus connected apps like Google Workspace, and writes a cited report.
• Web search is for quick answers; Research is for questions with many angles.
• Good prompts state the decision, the scope and the sources. Good habits check the citations behind your key claims.
• Claude Code has an official /deep-research workflow that cross-checks claims in the background.

The catch with any research tool is what happens next week. You run the same kind of question again, retype the same brief from memory, and paste the result into the same doc. Taku is an AI-native desktop workspace where you can save a research workflow that worked and run it again instead of starting from a blank prompt. Taku is in Beta, and the Mac app is available now.