OpenAI Unveils Deep Research, a Highly Transparent, Thorough Model for Work-Ready Research Reports

OpenAI is unveiling a new agentic capability for ChatGPT designed to streamline complex research processes, turning over in-depth, comprehensive responses to research queries in as little as tens of minutes. Developed for entities in intensive knowledge work areas—such as finance, science, policy, and engineering—deep research synthesizes hundreds of online resources to produce precise, reliable research.

Deep research independently discovers, reasons about, and consolidates insights across high quantities of internet resources—such as text, images, and PDFs. Executing multi-step research for complex tasks with ease, deep research aims to deliver the power of a research analyst with little effort on behalf of the user.

Responses delivered with deep research are accompanied by clear citations and a summary of its reasoning, establishing trust through response transparency. Information surfaced can be easily referenced and verified, making it invaluable for the purpose of locating non-intuitive information that would only be discoverable by scouring multiple websites, according to OpenAI.

Deep research is powered by the same reinforcement learning methods behind OpenAI o1, OpenAI’s first reasoning model. This capability expands on what OpenAI o1 could achieve, merging the power of technical expertise with extensive context and information gathering from a variety of online resources.

In ChatGPT, users can select “deep research” in the message composer and begin entering queries. Returning a comprehensive response can take anywhere between 5 to 30 minutes; users receive a notification when research is completed. Responses returned by ChatGPT are delivered as a report within the chat, which will soon offer embedded images, data visualizations, and other analytical features.

Utilizing deep research is the difference between generating a quick summary and a highly transparent, thorough response that can be used as a work product, according to OpenAI. According to an evaluation from Humanity's Last Exam, the deep research model achieved a new high in accuracy at 26.6%.

Notably, OpenAI emphasizes that deep research is still capable of hallucinating facts or inferring incorrectly—though at a lower rate than existing ChatGPT models. OpenAI anticipates that these challenges—including distinguishing authoritative information from a rumor and failing to illustrate uncertainty accurately—will improve as the model is used.

With its power, Deep research in ChatGPT is currently very compute intensive, according to OpenAI. Paid users will soon receive significantly higher rate limits as OpenAI develops a faster, more cost-effective version of deep research. The model is currently available to Pro users, which will be followed by Plus and Teams users, and finally Enterprise.

To learn more about deep research, please visit https://openai.com/.