Deep research

Deep research is a longer job than an ordinary chat reply: make a plan, search, read, change course, and write a report you can check. Google shipped it in the Gemini app on 11 December 2024, for Gemini Advanced subscribers. You approve a multi-step plan, Gemini browses for a few minutes, and you get a report with links to the pages it used, which you can export to a Google Doc. Google called it the first agentic feature in Gemini¹.

OpenAI launched its version in ChatGPT on 2 February 2025: an agent that finds, analyses and synthesises hundreds of online sources into a cited report. At launch it ran on a version of o3 tuned for browsing, and a task could take 5 to 30 minutes. OpenAI said it could still state false facts or draw a wrong inference, less often than other ChatGPT models in internal tests, and that it often failed to show when it was unsure².

OpenAI's help centre, current in October 2026, says you start with /Deepresearch or from the tools menu, review the plan, and interrupt the run to change the question or the sources. It can use the public web, files you upload, and connected apps that are allowed to read. It does not use those apps to change your files. The report has a contents list, sources, and a history of the steps, and you can download it. How many runs you get depends on the plan. A counter shows what is left, and a fixed monthly allowance resets 30 days after first use. It is not offered in every country. Use ordinary search for a quick fact, and deep research when the question needs many sources³.

Google's developer docs describe the current agent as planning, searching and writing a cited report over several minutes, in the background rather than in one chat reply⁴. Perplexity's help centre calls its version Research mode. It runs dozens of searches, reads hundreds of sources, and Perplexity says most tasks finish in under three minutes, though a reply can take four or five. Free accounts get limited use, and you cannot pick the model⁵.