On 10 February 2026, OpenAI moved Deep Research in ChatGPT to the GPT-5.2 model and reworked the mode itself. Before it starts, it now shows a research plan that you can edit; while it works, you can follow its progress and steer it; and you can limit its search to sites you have chosen or to connected apps. The result opens as a full-screen report with a table of contents and a list of sources, and it can be exported to PDF, Word or Markdown.
What changed in February
Previously, Deep Research ran on a special version of the o3 model and, once the limit was used up, switched to the lighter o4-mini. Even then it asked clarifying questions before starting, and from November 2025 you could interrupt it and add to the task. Since the update:
- the plan can be edited before the start: drop the lines of inquiry you don’t need, add the ones you do, change the order;
- you can follow the research in real time and change its focus and set of sources along the way;
- you choose the sources: only the sites you specify, or priority for them while still searching the wider web. Apps and MCP servers could be plugged in through connectors from mid-2025; now the range is wider;
- the report has become a document: a separate full-screen view, a table of contents, a section listing the sources used, an activity history and export.
OpenAI itself did not name the model in its official texts: the company wrote about GPT-5.2 on its X account. The old Deep Research mode was removed on 26 March 2026; chat history was kept.
How to get the most out of the mode
Deep Research works best when the request is phrased as the document you expect. Not “what’s new there”, but “put together an overview with a timeline, the key decision points and conclusions”. Not “how does this work”, but “explain the mechanism, compare the approaches and point out where we might be wrong”.
Set boundaries: the period, the geography, the type of sources and the format of the result. Use plan editing — it is cheaper than redoing a finished report. And after the research, check the key numbers and links separately: the mode assembles text faster than a human, but it makes mistakes just as confidently as any language model. The more important the decision, the less you should rely on a single automatically assembled document.
Limits
OpenAI does not currently publish exact numbers. On the pricing page, Deep Research is “limited” on the Free and Go plans, “expanded” on Plus and “maximum” on Pro. The remaining allowance is shown in a counter inside the product, and the monthly limit resets 30 days after the first run.
What happened next
GPT-5.2 itself was removed from ChatGPT on 12 June 2026. As of September, OpenAI’s help page says that Deep Research runs “on the latest models” by default and that you can choose the previous one if you wish. OpenAI does not specify which model is inside the mode now: there was no separate announcement of a move to GPT-5.5, 5.6, GPT-6 or GPT-6.1, released in the autumn.
What changed in the model line-up over these months is covered in the breakdown of GPT-5.2 and what came after. For what else was going on around the model in those weeks, see the big AI digest.
