OpenAI has updated Deep Research in ChatGPT, moving the feature to the GPT-5.2 model and adding new controls for how research is run and reviewed. The upgrade expands what users can target during a research session, including the ability to connect supported apps and focus on specific websites. OpenAI is positioning the change as a practical step toward more controllable, agent-like research inside ChatGPT.
What Changed With the GPT-5.2 Update
Deep Research now runs on GPT-5.2, replacing earlier model backends used for the feature. The update also introduces more flexible research controls, letting users narrow where the system looks and how it proceeds through a multi-step search process.
In addition, results can be presented as full-screen reports, designed to make longer outputs easier to read and navigate.
New Controls: Apps, Site Focus, and Live Progress
A key addition is the option to connect apps to ChatGPT and incorporate those sources into Deep Research. Users can also direct the system to search specific websites, which can be useful when the goal is to rely on known repositories, official documentation, or a defined set of publishers.
The research process is now more interactive: progress can be tracked in real time, interrupted with follow-up questions, and supplemented with additional sources while the session is running. This shifts Deep Research from a single, locked workflow into a more guided process where the user can steer scope and inputs midstream.
Why OpenAI Calls It an “AI Agent”
OpenAI has described Deep Research as its first “AI agent” inside ChatGPT because it can independently initiate multi-stage web research based on a user request before producing a synthesized response. Deep Research originally launched in 2025 and previously ran on models including o3 and o4-mini.
Reliability: Better Search, Same Need for Verification
Even with improved browsing and more targeted controls, web-enabled research does not fully eliminate generative AI errors. Longer outputs still carry a higher risk of mistakes, and accuracy depends on both the quality of sources and how the question is framed.
In everyday use, narrowing the search to specific sites and using focused queries can improve reliability, but users should still verify key claims, especially for high-stakes topics.
Conclusion
The GPT-5.2 upgrade pushes Deep Research toward a more controllable research workflow, adding app connections, site-specific searching, and live, interruptible progress. The full-screen report format aims to make long research outputs easier to consume. While these changes can reduce uncertainty and improve sourcing, careful verification remains essential because browsing lowers error rates but does not eliminate them.
