Deep research (ChatGPT)

What it is

Deep research is a research agent inside ChatGPT. OpenAI released it on 2 February 2025, built on a version of the o3 model.^[1][4] Rather than answering a question in a single reply, it goes out to the web, gathers material and summarises it into a written report.^[2] A report takes minutes to about half an hour to produce, so using it means waiting rather than watching an answer appear word by word.^[2][5]

The distinction matters. A chat model answers largely from what it learned during training, plus whatever text is pasted into the conversation. A research agent is given tools to search and open pages and a budget of steps in which to use them. The output is a structured document — findings, sections, sources — instead of a paragraph, and the user's job shifts from asking follow-ups to reading and checking a draft.

How OpenAI framed it

OpenAI described the tool as compressing work that would take a human many hours into tens of minutes, synthesising hundreds of online sources into a report at the level of a research analyst.^[5] That framing set the expectation: not a chat answer with a few links attached, but a multi-source document someone could build on. Whether a given report reaches that level depends on the question, which is why the limitations below matter as much as the pitch.

Access and limits

At launch it was offered as a version optimised for Pro users, with up to 100 queries a month; Plus and Team were to follow.^[5] On 24 April 2025 OpenAI added a lighter version powered by o4-mini and set monthly limits: 25 queries for Plus, Team, Enterprise and Edu, 250 for Pro and 5 for Free.^[5] Metering queries by plan says something about cost: a single report can involve many searches and a long generation, so access is rationed rather than unlimited.

Known weaknesses

OpenAI acknowledged that the agent could still hallucinate facts, struggle to separate authoritative information from rumour, and misjudge how uncertain it should sound.^[5] These are familiar problems of language models, but the format makes them easier to miss: a report that reads smoothly and looks finished invites trust. Checking sources and going back to the original documents for key figures stays with the reader.

Where it sits

In July 2025 the ability to gather and summarise information was inherited by ChatGPT agent.^[2] The API versions released with it, o3-deep-research and o4-mini-deep-research, were deprecated on 22 April 2026 and shut down on 23 July 2026.^[3] Research first and write afterwards, rather than ask and answer, became part of how ChatGPT is used.