ChatGPT Deep Research is the mode that spends 5 to 30 minutes browsing, reading and writing a cited report for you. It is one of the most useful tools in ChatGPT, and one of the most misunderstood.
This guide shows how to start a run, how limits work on each plan in 2026, how to use connectors and site limits, and when another mode is the better pick.
- What it is: an agent inside ChatGPT that researches for 5 to 30 minutes and returns a structured report with citations.
- Start it: open the tools menu (the + in the composer), pick Deep research, then review the research plan.
- Limits: OpenAI now uses labels, not numbers: Limited (Free, Go), Expanded (Plus), Maximum (Pro). The counter resets 30 days after your first use.
- Since Feb 2026: connected apps, site limits, live progress, mid-run follow-ups and a full-screen report with PDF, Word and Markdown export.
- Best for: complex, multi-source questions. Use normal chat for quick facts.
What is ChatGPT Deep Research?
Deep Research is an agent built into ChatGPT. You give it a question, it plans the research, searches and reads many sources, then writes a report with citations.
It launched in February 2025 on a special version of o3. In February 2026 OpenAI moved it to a model based on GPT-5.2 and added better controls.
What it can read
- The web: public pages, which it browses on its own.
- Your files: PDFs, spreadsheets and documents you attach.
- Connected apps: sources like Google Drive, Slack or Notion, in read-only mode.
- Chosen sites: you can limit it to specific domains you trust.
Deep Research limits per plan (2026)
In 2025 OpenAI published counts like 25 tasks a month on Plus. It has since dropped the numbers. The pricing page now uses labels only.
| Plan | Price (US) | Deep Research |
|---|---|---|
| Free | $0 | Limited |
| Go | $8 per month | Limited |
| Plus | $20 per month | Expanded |
| Pro | From $100 per month | Maximum |
| Business | $25 per seat monthly, $20 yearly | Included, extra runs use workspace credits |
| Enterprise | Custom | Included, admins control it by role |
| Edu | Custom | Included, reported as 5 tasks per rolling 24 hours, then credits |
How the counter resets
- In-product counter: ChatGPT shows how many runs you have left.
- 30-day window: fixed allowances reset 30 days after your first use, not on the 1st of the month.
- Credits: on Business and credit-based Enterprise, a task is reported to cost 50 credits after the included use.
Note: OpenAI’s help center blocks automated readers. The labels above come from the live ChatGPT pricing page, and the reset and credit details come from reports that quote OpenAI’s help pages. Your counter in ChatGPT is the final word.
For a full plan breakdown, see ChatGPT Free vs Go vs Plus vs Pro.
How to use Deep Research, step by step
- Open a new chat. Deep Research works on web, desktop and mobile.
- Pick the tool. Select the + (tools) button in the composer, then choose Deep research.
- Add sources if you want. Attach files, turn on connected apps, or limit it to specific sites.
- Write a clear question. Say the goal, the audience, the time range and the format you want.
- Review the research plan. ChatGPT shows its plan first. Edit it to narrow scope, add sources or change the angle.
- Let it run. Watch live progress, or leave. ChatGPT notifies you when the report is ready.
- Steer mid-run. Add a follow-up or a new source while it works, without starting over.
- Open the report. Read it in the full-screen viewer, check citations, then export.
Tip: Editing the research plan is the single best way to improve a report. Two minutes there can save a wasted 20-minute run and a used-up credit.
Connectors, files and site limits
Since the February 2026 update, Deep Research can pull from more than the open web.
Connected apps
- What they add: your own documents, messages and notes as sources.
- Examples: Google Drive, Slack and Notion.
- Access: read-only. Deep Research does not change anything in your apps.
- Setup: connect apps in ChatGPT first, then pick them before the run.
Developers and admins can also add data through MCP servers. My best MCP servers list is a good starting point.
Limiting to specific sites
You can tell Deep Research to search only chosen sites, such as government pages, journals or an industry database. This is ideal for medical, legal or financial topics, where low-quality blogs cause trouble.
Warning: Connected app data can be sensitive. On work accounts, check your company policy before pointing Deep Research at shared drives or chat channels.
Reading and exporting the report
Reports open in a full-screen viewer, separate from the chat.
- Left column: a table of contents to jump between sections.
- Right panel: the list of sources and citations.
- Export: download as PDF, Word or Markdown.
- Follow-ups: ask questions about the report in the same chat.
Open two or three key citations before you share anything. It takes a minute and catches most errors.
Prompt tips that get better reports
A vague prompt gets a vague report. Give Deep Research the same brief you would give a human researcher.
Goal: Compare the top 4 project management tools for a 10-person agency.
Audience: non-technical founder.
Scope: pricing, client portals, time tracking. Data from 2026 only.
Sources: official pricing pages first, then reviews.
Format: summary table, then one section per tool, then a recommendation.
- Name the time range: say “2026 only” so it skips old data.
- Rank sources: ask for official pages first.
- Ask for a table: comparisons read better in a table.
- Ask it to flag doubt: “mark any claim with only one source”.
- Split big topics: two focused runs beat one huge one.
Good use cases for Deep Research
Deep Research shines when a question has many moving parts. Here are jobs where it saves real time.
- Buying decisions: compare software, laptops or suppliers across pricing pages and reviews in one report.
- Market scans: list the main players in a niche, with funding, pricing and features.
- Literature reviews: summarize recent studies on a topic, limited to journal sites.
- Policy checks: gather official rules from government sites for a country or state.
- Trip planning: build a detailed plan with visa rules, transport options and costs.
- Internal digests: pull notes from connected Drive files or Slack channels into one summary.
When not to use it
- Simple facts: a normal chat answers in seconds and does not touch your quota.
- Breaking news: the last few hours may not be indexed yet, so search mode is faster.
- Tasks that need actions: booking, filling forms or editing files belong in ChatGPT Work.
Saving runs for the big questions keeps your monthly allowance for when it really matters.
Deep Research vs other ChatGPT modes
ChatGPT now has several ways to work, and each has a job.
| Mode | Time | Output | Best for |
|---|---|---|---|
| Normal chat with search | Seconds | Short answer with links | Quick facts and news |
| Thinking models | Seconds to minutes | Reasoned answer | Math, code, logic |
| Deep Research | 5 to 30 minutes | Cited report | Multi-source research |
| ChatGPT Work / agent | Minutes to hours | Finished task | Multi-step actions, files |
| Dots (Pro, Business Premium) | Ongoing | Background work | Always-on jobs |
Dots are covered in my ChatGPT dots guide. If you also use Google tools, compare it with Gemini Deep Research.
Accuracy and safe use
OpenAI says Deep Research can still hallucinate facts, repeat rumors or state things with too much confidence. A cited report is not the same as a checked report.
- Check key numbers: open the cited page for every figure you plan to reuse.
- Watch dates: old pages can rank well. Confirm the date on each source.
- Avoid big decisions on one run: medical, legal and money choices need an expert.
Troubleshooting: Common Problems and Fixes
- No Deep research option: it depends on country and plan. On Enterprise or Edu, your admin may have turned it off for your role.
- “Limit reached”: wait for the 30-day reset shown in the counter, upgrade, or use credits on business plans.
- The run seems stuck: big runs can take 30 minutes. Check live progress before you cancel.
- The report is too broad: edit the research plan next time and add scope, dates and format.
- A connected app returns nothing: reconnect it in ChatGPT and confirm you selected it before the run.
- Weak sources: limit the run to specific trusted sites.
- ChatGPT will not load: see ChatGPT not working fixes.
Frequently Asked Questions
How many Deep Research runs do I get on Plus?
OpenAI no longer publishes a number. Plus is labeled “Expanded”, and the in-product counter shows what you have left.
Is Deep Research free?
Free and Go users get limited access. Plus, Pro and business plans get more.
How long does a Deep Research report take?
Usually 5 to 30 minutes, depending on the topic. ChatGPT notifies you when it is done.
When do Deep Research limits reset?
Fixed allowances reset 30 days after your first use in the current window, not on a calendar date.
Can Deep Research read my Google Drive?
Yes, if you connect it. Connected apps are used read-only as sources.
Can I export a Deep Research report?
Yes. Reports can be exported as PDF, Word or Markdown from the full-screen viewer.
What model powers Deep Research?
Since February 2026 it uses a model based on GPT-5.2. It started on a version of o3 in 2025.




