OpenAI launched Deep Research on February 2, 2025, with a system card dated February 25 describing the system in detail. It is an agentic capability inside ChatGPT that conducts multi-step research on the internet for complex tasks, powered by an early version of the o3 reasoning model optimized for web browsing. Given a prompt, it searches, clicks, scrolls, reads files and PDFs, and writes and runs Python in a sandbox to analyze data, pivoting its approach as it encounters new information - and then synthesizes a comprehensive, cited report at the level of a research analyst.
The model was trained with reinforcement learning on browsing tasks, learning the core skills of searching, clicking, scrolling, interpreting files, and reasoning across many websites to find specific facts or write reports. On Humanity’s Last Exam, an exam-style benchmark designed to be extremely hard, OpenAI reported a then-record 26.6 percent. The system card details the new safety work browsing introduced: stronger protections around personal information published online, and training the model to resist malicious instructions it might encounter on web pages.
Deep Research arrived weeks after Google’s Gemini Deep Research, and the two together established the research agent as a distinct product category. It also became an early flagship for what the o-series reasoning models could do when given tools and a long time horizon rather than a single-shot answer.