OpenAI has launched "deep research" in ChatGPT, a new agentic capability powered by an optimized version of the o3 model that conducts multi-step internet research to produce comprehensive reports. This feature is designed for intensive knowledge work, allowing users to offload complex web research tasks that would typically take hours into queries completed in tens of minutes.
The system leverages reasoning to search, interpret, and analyze massive amounts of text, images, and PDFs, synthesizing hundreds of online sources with clear citations. It was trained using end-to-end reinforcement learning on hard browsing and reasoning tasks, enabling it to plan multi-step trajectories and react to real-time information.
On the Humanity’s Last Exam benchmark, the model achieves 26.6% accuracy across over 3,000 expert-level questions. It also reaches a new state of the art on the GAIA benchmark for real-world questions requiring reasoning and tool-use proficiency.
Initial access is restricted to Pro users in specific regions, with Plus, Team, Enterprise, and Edu users receiving increased monthly query limits via a lightweight version powered by o4-mini.