OpenAI launches deep research in ChatGPT

Background

At the end of 2024, OpenAI announced a reasoning model called o3 but had not yet released it^[1]. Reasoning models differ from earlier chat models in how they use computation: rather than producing an answer as soon as the next piece of text is generated, they work through intermediate steps before replying, and a system allowed to run longer can chain more of those steps together.

What happened

On 2 February 2025, OpenAI launched a feature called deep research inside ChatGPT^[1][3]. It was offered first to ChatGPT Pro subscribers, on the day of launch^[3]. The company said limited access would be extended to ChatGPT Plus, Team and Enterprise users in the weeks that followed^[3].

How it worked

Deep research ran on a version of the o3 reasoning model^[1]. The feature therefore reached users on a model that OpenAI had announced but not released on its own^[1]. Instead of answering within seconds, it worked as an agent: the user describes what should be investigated, and the system carries out the steps and returns a research report — a structured document setting out a question and what was found — rather than a short chat reply. Each report took between 3 and 30 minutes^[2]. OpenAI's own description of the launch gave a range of 5 to 30 minutes, and said ChatGPT could take up to 30 minutes to generate a comprehensive deep research response^[3][4].

Agents of this kind are commonly built around a loop: plan the next step, take it, look at the result, decide what to do next. Running such a loop for tens of minutes is what the long runtimes reflect. The trade-off falls on the user's side: the request is handed over and the answer arrives later, to be read and checked, instead of being steered turn by turn.

What came next

OpenAI later combined the capabilities of deep research with Operator in ChatGPT agent^[2]. The research workflow became one part of a broader agent product rather than a feature standing alone.

Why it is worth knowing

Deep research is a marker of how chat assistants changed shape. The interface was still a text box, but the thing behind it was no longer only a responder; it took on a task and came back with a document. That shift also made the release schedule of the underlying models visible in an unusual way: the public could use a feature built on a reasoning model that had been announced but not yet shipped by itself^[1]. And the staged rollout, from Pro users on launch day to wider access afterwards^[3], shows the order in which such features commonly arrive — the highest tier first, more users later.