What is ChatGPT deep research, and how does an AI research agent work?

Yesterday, Sunday, February 2, 2025, OpenAI launched deep research, a new mode in ChatGPT built for questions that need more than a quick answer. You describe what you want to know, and ChatGPT goes off on its own for anywhere from 5 to 30 minutes, searching the web, reading what it finds, changing course when it learns something new, and finally handing back a structured report with sources for each claim.

OpenAI calls it an agent, an AI that carries out a multi-step task by itself. It is the company's next agent after Operator, which takes actions on websites for you. Where Operator clicks buttons, deep research reads and writes. OpenAI says it finishes in minutes research that would cost a person many hours.

What OpenAI launched

Deep research is aimed at people who do intensive knowledge work, in finance, science, policy or engineering, but OpenAI also pitches it to shoppers comparing cars, appliances or furniture. In practice it works like this:

1. In ChatGPT, you select deep research in the message box and type your question. You can attach files or spreadsheets for context.
2. A sidebar opens, showing a summary of the steps it is taking and the sources it is using.
3. You can leave and do something else. ChatGPT sends a notification when the report is ready.
4. The report arrives in the chat, with citations and a summary of the model's thinking. For now it is text only; OpenAI plans to add images and charts.

Access is limited at the start. Subscribers to the ChatGPT Pro plan get up to 100 requests a month. Plus and Team users come next, with OpenAI aiming for Plus in about a month, and then Enterprise. It is not yet available in the United Kingdom, Switzerland or the European Economic Area, and it works on the web version of ChatGPT, with the mobile and desktop apps to follow this month.

How it works

Deep research runs on a version of o3, OpenAI's upcoming reasoning model, adapted for browsing the web and analyzing data. A reasoning model thinks through a problem step by step before answering, rather than replying immediately.

OpenAI trained it with reinforcement learning, a method in which a model learns by trial and error and is rewarded when it gets closer to the goal. Here the practice tasks were hard, real-world research problems that required using a web browser and Python, a programming language. Through that training, the model learned to plan a sequence of searches, back up when a path leads nowhere, and react to what it finds.

A useful comparison is a research assistant in a library. A regular chatbot answers from memory, like someone who has read a lot but cannot check. ChatGPT's search feature looks things up and summarizes them quickly. Deep research is the assistant who spends the afternoon in the stacks: it reads dozens of sources, follows leads, runs calculations, draws graphs with Python and notes exactly which sentence supports each point.

OpenAI tested it on two public benchmarks, standard sets of test questions. On Humanity's Last Exam, more than 3,000 expert-level questions across over 100 subjects, deliberately designed to be very hard, it scored much higher than other models:

On GAIA, a benchmark of real-world questions that require browsing and tool use, it averages 67.4% on a first attempt, above the previous best of 63.6%. OpenAI notes that some GAIA answers had leaked online, and that it blocked those pages during testing.

Limits to keep in mind

OpenAI lists the weaknesses plainly. Deep research can still hallucinate, meaning state things that are not true, or draw wrong conclusions, though OpenAI's internal tests show it does so less often than existing ChatGPT models. It can struggle to tell an authoritative source from a rumor. It often fails to signal when it is unsure. And at launch, reports and citations may have formatting errors.

TechCrunch raises the human side of the question: a long, well-cited report looks convincing, and the real test is whether readers will check the sources or simply copy the text. TechCrunch also notes that Google announced a similar feature, with the same name, less than two months ago.

Why it matters

Deep research changes the unit of work you can hand to an AI. Instead of one question and one answer, you can delegate a small research project and come back to a draft. For students, analysts and anyone who compares options before a big decision, that could save hours, provided the citations are checked.

It also shows where OpenAI is heading. The company says it expects to connect deep research to specialized and subscription data sources, and imagines it working together with Operator: one agent that investigates, another that acts.

How do you write a good request for deep research?

Because each request takes many minutes and counts against a monthly limit, it pays to be precise. State the question, who the report is for, the scope (countries, years, price range), the sources you trust or distrust, and the format you want. If you have data, attach it.

I am choosing an electric bike for a 12 km daily commute with hills,
budget under $2,500, sold in Canada. Compare five current models on range,
weight, motor type, warranty and price. Prefer manufacturer specs and
independent reviews; avoid affiliate "top 10" lists. Give me a table,
then a short recommendation, and cite a source for every number.

When the report arrives, read the sidebar to see how it got there, open the citations for the claims that matter most, and ask follow-up questions in the same chat.