Deep research is a ChatGPT feature that takes a question, makes a research plan, browses many web pages (and files you give it) for several minutes, and returns a report with citations. It launched on February 2, 2025, running on a version of OpenAI’s o3 model tuned for web browsing and data analysis. It is not the same as extended thinking, which makes a model reason longer inside a single reply without a research workflow. Details below are from OpenAI’s pages as of October 2026; plans, limits and model names change often, so check the linked pages before relying on a number.
What “o3” means here
o3 is a model name, a member of OpenAI’s “o-series” of reasoning models. It is not an acronym and it is not a technique called “objective-oriented optimization”, despite what some articles (including an earlier version of this one) claimed. Reasoning models are trained to work through a problem step by step before answering. Deep research took such a model and trained it further for multi-step browsing and analysis tasks, so it can search, read pages, change course based on what it finds, and write up the result. OpenAI’s announcement describes it as using reasoning to search, interpret and analyze large amounts of text, images and PDFs.
OpenAI has not published the network design or the training details beyond that, so descriptions of its internal architecture (transformers plus graph networks and so on) are guesses. Skip any article that states them as fact.
How deep research works in ChatGPT
OpenAI’s help center describes the flow like this:
- State the outcome you want: a question, a comparison, a market scan, a literature overview.
- Choose sources. By default it can use the public web and files you upload. You can restrict it to particular sites, or prefer them while still searching more widely.
- Review the plan. ChatGPT proposes a research plan first, and you can edit it before it starts.
- Let it run. You can watch progress, interrupt, and refine the focus or sources mid-way.
- Read the report. It opens full-screen with a table of contents, a sources section and an activity history, and every output includes citations or source links. You can download reports in formats such as Markdown, Word and PDF.
OpenAI’s launch material says a task can take roughly 5 to 30 minutes, and you get a notification when it is done. The “30 minutes” in older headlines is the top of that range, not a fixed thinking period.
Deep research vs extended thinking
People search for these two together because both are “ChatGPT thinking longer”. They solve different problems.
| Deep research | Extended thinking | |
|---|---|---|
| What it is | An agent that plans, searches, reads and writes a report | A setting that gives a reasoning model more time to think before one answer |
| Sources | Many web pages and your files, with citations | Mainly the model’s own knowledge plus whatever tools the chat has |
| Typical time | Minutes (OpenAI cites 5 to 30) | Seconds to a few minutes, depending on the model and setting |
| Output | A structured, cited report | A normal chat answer |
| You steer it | Edit the plan, interrupt, restrict sources | Choose the thinking level, if your plan offers it |
| Best for | “Find and compare what’s known about X” | Hard reasoning: math, logic, debugging, tricky code |
Rule of thumb: if the answer depends on gathering information from many places, use deep research. If the answer depends on thinking hard about information you already have, use a thinking model. The exact thinking levels (Standard, Extended and others) differ by plan and model and have been renamed more than once, so look at the model picker in your own account rather than trusting a screenshot.
Access and limits
Availability depends on your plan and region, and OpenAI adjusts quotas over time. Reported figures have ranged widely since launch (for example 100 queries a month for Pro at the start, later raised), and current OpenAI documentation describes usage generally as “varies by plan”, with ChatGPT showing how many tasks you have left. Chat and Work have separate allowances. No number is quoted here on purpose; check OpenAI’s deep research help article and your own account.
Using it through the API
Developers can use deep research through the Responses API. OpenAI’s guide lists the models o3-deep-research and o4-mini-deep-research. As of October 2026 the guide flags both for retirement with a replacement named on that page, so read it for current model ids before building anything. What stays the same in the guide:
- You must give at least one data source: web search, a remote MCP server, or file search over vector stores. Code interpreter is optional.
- Function calling is not supported.
- Runs are long (“tens of minutes”), so use background mode or a very high client timeout.
max_tool_callscaps the number of searches and tool calls, which is the main control over cost and latency.- The API does not ask clarifying questions, so write a fully specified prompt.
- The response lists the actions taken and ends with a message containing inline citations.
The same guide warns about prompt injection: a web page or file can contain instructions aimed at the model. Connect only trusted MCP servers and files, and log tool calls. Read the official deep research guide before sending it private data.
Where it works well, and where it does not
Good fits: a first overview of a field, comparing products or libraries, collecting sources on a policy or technology question, summarizing documents you uploaded, preparing background before a technical decision. Poor fits: anything where the answer is a single fact (use normal search), problems that need original proof or experiments, and anything where a wrong detail is costly and nobody will check it.
OpenAI itself lists limitations. At launch it said the feature can still hallucinate facts or make wrong inferences (at a lower rate than earlier ChatGPT models, by its internal evaluations), can struggle to tell authoritative sources from rumors, and is often poorly calibrated about its own uncertainty. Reports can have formatting errors in citations. Treat the output as a well-sourced draft:
- Open the cited links for every claim you will repeat.
- Check dates, prices and version numbers on the vendor’s own page.
- Be suspicious of confident numbers with no source.
Getting better reports
- Say what the report is for. “Compare three message queues for a team of five running on AWS” beats “tell me about message queues”.
- Name the format. A table of options with a recommendation, or a one-page brief, saves a rewrite.
- Restrict sources when quality matters: official docs, specific sites.
- Fix the plan. Remove sub-questions you do not need before it starts; it is cheaper than a long run in the wrong direction.
- Attach your own files when the answer must reflect your data.
- Follow up in the chat to narrow a result instead of starting over.
What it means for learning to code
Deep research is useful for background work: comparing frameworks, reading up on a protocol, summarizing a library’s docs. It is a poor way to learn programming itself, because reading a polished report is not the same as writing and debugging code. If you use AI while learning, use it the way the posts on learning to code with AI tutors and the best AI for coding suggest: ask for hints and explanations, then write the solution yourself.
Frequently asked questions
What is OpenAI deep research?
A ChatGPT feature that plans a research task, browses many sources for several minutes, and returns a cited report.
Does deep research use o3?
It launched on a version of o3 tuned for browsing and analysis. Model names behind ChatGPT features change, so check OpenAI’s help page for what your account uses today.
How long does it take?
OpenAI’s launch material says about 5 to 30 minutes, depending on the task.
Is deep research the same as extended thinking?
No. Extended thinking gives a model more time to reason about one answer. Deep research is a multi-step workflow that gathers sources and writes a report.
Can I trust the report?
Not without checking. It cites sources, but OpenAI acknowledges it can still get facts or inferences wrong, so verify anything important.
