OpenAI Introduces New ‘Deep Research’ Agent on ChatGPT: For Multi-Step Research Tasks
A new agent has now been introduced by OpenAI with the capability of conducting multi-step research and generating comprehensive reports at a level comparable to a research analyst. The feature is already made available to a few users and more will be getting it soon.
Read more about it below.
Deep Research for Multi-Step Research Tasks
As mentioned above, Deep Research is a new agentic capability that OpenAI has introduced on ChatGPT and it will be able to conduct multi-step research to fulfill complex tasks that would probably require hours for a human in just a matter of minutes. With the right prompt, Deep Research on ChatGPT will move forward by finding, analyzing, and synthesizing plenty of online resources to provide a comprehensive report of the assigned task. It is powered by the web browsing and data analysis optimized web version of OpenAI o3 model, and is particularly made for people doing work requiring intensive knowledge. To add, the new agent or tool was trained based on real world tasks with Python and browser, along with the same reinforcement learning methods as OpenAI o1.
On how to use it, users will be able to find it on the message composer, select it and enter the query. Attaching files and spreadsheets are also possible. Once it starts working, it may take from five to thirty minutes to get it completed, and a sidebar will be displayed showing the steps and sources. At the end, a comprehensive report will appear in the chat. Also, in the coming weeks, embedded images, data visualizations, and other analytic outputs will also be added to the Deep Research agent, bringing more clarity and context.
Speaking more about its test results, on Humanity’s Last Exam, Deep Research acquired 26.6% accuracy. On the GAIA public benchmark, it topped the external leader board. It also automated many hours of difficult and manual investigation in an internal test evaluation across a range of expert level tasks. However, this agent or tool is still in its early stages and has a few limitations which includes hallucinating facts in responses and making incorrect inferences. But this is at a lower rate in comparison with other ChatGPT models that are currently available to users.
As of now, Deep Research is already made available to Pro users with 100 queries per month. Plus and Team users will get it next and later for the Enterprise users. Access to users in the UK, Switzerland, and Europe are also being worked on. To note, it is only available on ChatGPT web currently, and within a month, it will also be brought to the mobile and desktop apps too.
Stay tuned.
