OpenAI Launches Deep Research, an Agent That Researches for 30 Minutes

OpenAI unveiled deep research this Sunday, a new ChatGPT capability that functions as a research agent: instead of firing back a quick answer, it spends between five and thirty minutes browsing the web, cross-referencing sources, and drafting a documented report complete with citations. The company describes it as a tool for "people who do intensive knowledge work in areas like finance, science, policy, and engineering and need thorough, precise, and reliable research."

The idea breaks with the chatbot's usual logic. Until now, ChatGPT rewarded speed: you asked, it answered in seconds. Deep research bets on the opposite. Users agree to wait half an hour in exchange for a document that has combed through multiple websites and other sources before writing a single line.

What it is and how it works

To activate it, users select the "deep research" option in ChatGPT's composer and enter their query, with the option to attach files or spreadsheets. For now it's a web-only experience; OpenAI plans to bring it to the mobile and desktop apps later this month. Once the search wraps up, the user gets a notification.

For now, outputs are text-only. OpenAI has said it intends to soon add embedded images, data visualizations, and other "analytic" outputs. Also on its roadmap is connecting to "more specialized data sources," including "subscription-based" resources and companies' internal data.

The rollout starts today for ChatGPT Pro users, capped at 100 queries per month. It will then reach Plus and Team plans, and finally Enterprise. OpenAI is targeting a Plus rollout in about a month, and expects query limits for paid users to rise "significantly" soon. The launch is geo-targeted: the company gave no timeline for customers in the U.K., Switzerland, and the European Economic Area.

The engine: a special version of o3

Accuracy is the big unknown for any tool marketed as "research." Generative AI hallucinates — it invents plausible-looking data — and makes mistakes that, in a research context, can be especially damaging. OpenAI is tackling this on two fronts.

The first is traceability. Every deep research output will, according to the company, be "fully documented, with clear citations and a summary of [the] thinking, making it easy to reference and verify the information." The goal is for readers to be able to check where every claim comes from.

The second is the underlying model. Deep research runs on a special version of o3, the "reasoning" model OpenAI recently introduced, trained through reinforcement learning on "real-world tasks requiring browser and Python tool use." Reinforcement learning trains the model through trial and error: the closer it gets to a goal, the more virtual "rewards" it receives, which in theory sharpens it for that task.

OpenAI says this version of o3 is "optimized for web browsing and data analysis," adding that "it leverages reasoning to search, interpret, and analyze massive amounts of text, images, and PDFs on the internet, pivoting as needed in reaction to information it encounters." The model can also browse files uploaded by the user, generate and refine charts with Python, embed those charts and images from websites into its responses, and cite specific sentences or passages from its sources.

The eye-catching number: Humanity's Last Exam

OpenAI tested deep research on Humanity's Last Exam, a benchmark with more than 3,000 expert-level questions across various academic fields. The o3 model powering the tool scored 26.6%.

26.6% might look like a failing grade, but context changes everything: Humanity's Last Exam was specifically designed to be tougher than other benchmarks so it stays ahead of model advancements. Compared to its rivals, the result is decisive. According to OpenAI, the model came in far ahead of Gemini Thinking (6.2%), Grok-2 (3.8%), and its own GPT-4o (3.3%). Against Gemini's 6.2%, deep research's accuracy is more than four times higher; against GPT-4o, it's eight times higher.

The limitations OpenAI itself acknowledges

The company isn't selling infallibility. It admits deep research sometimes makes mistakes and draws incorrect inferences. It can struggle to distinguish authoritative information from rumors, and often fails to convey when it's uncertain about something. It can also make formatting errors in reports and citations.

That last limitation — not knowing how to communicate its own uncertainty — is the trickiest one. A report with citations and a polished look conveys authority even when it contains errors, giving users less incentive to question something that appears well-sourced.

The track record calls for caution. ChatGPT's web search feature, ChatGPT Search, makes mistakes fairly often and gives wrong answers. TechCrunch's testing found it produced less useful results than Google Search for certain queries. Whether citations and traceability are enough to contain those errors when the volume of information processed is far greater remains to be seen.

A new category, and a recycled name

What matters about deep research isn't just what it does, but the category it opens up: an agent that works autonomously for minutes at a time, plans, reacts to what it finds, and delivers a finished product. It's a step beyond the conversational assistant that answers and waits for the next question.

That output format — lengthy, with cited sources — is, on paper, more appealing than the simple, reference-free summary of a conventional chatbot, especially for anyone worried about generative AI's impact on students or on people searching for information. But its value depends on how it's used: it remains to be seen whether most users will actually subject the report to real analysis and verification, or simply treat it as more professional-looking text to copy and paste.

One detail takes some shine off the launch: the name. Google announced a very similar AI feature, with the exact same name, less than two months earlier. The coincidence shows just how much the major labs are converging on the same bet — agents that research on their own — almost simultaneously.

What's left are the tests that really matter: the rollout to the most common paid plans, expanded query limits, the addition of charts and specialized sources, and the launch in Europe, which for now has no date. And, above all, whether an agent that spends half an hour researching produces reports that can truly be trusted.