Isla Fulford, a researcher at OpenAI, had a premonition that Deep Research would become popular even before its official launch. Fulford had a hand in creating the AI agent that autonomously navigates the internet, deciding which links to explore, what to read, and how to compile its findings into a comprehensive report. OpenAI initially rolled out Deep Research internally; whenever it experienced downtime, Fulford reported receiving a flood of inquiries from colleagues eager for its return. “The volume of DMs I received was quite encouraging,” remarks Fulford.
Since its public release on February 2, Deep Research has captured the attention of numerous users beyond the organization. “Deep Research has already produced 6 reports today,” Patrick Collison, CEO of Stripe, shared on X a few days post-launch. “It’s truly remarkable. Kudos to the team behind it.” “Deep Research is the AI tool that has genuinely engaged a significant portion of the policymaking community in DC, making them aware of AGI,” noted Dean Ball, a fellow at George Mason University specializing in AI policy.
Deep Research is included in the ChatGPT Pro subscription, priced at $200 per month. Users can provide queries like “Draft a report on the Massachusetts health insurance sector,” or “Summarize WIRED’s coverage of the Department of Government Efficiency,” and the model crafts a strategy, searching for pertinent websites, examining their content, and selectively choosing which links to pursue and which information warrants further exploration. After spending sometimes several minutes investigating, it synthesizes its findings into a thorough report that may also feature citations, data, and visual charts.
Currently, many tools labeled as AI agents function primarily as chatbots linked to basic programs without much complexity. The Deep Research model, however, engages in a form of artificial reasoning before formulating a plan and progressing through each stage. It also provides insights into the reasoning underlying its research in a companion window. “At times it’s like, ‘I need to reconsider; this path doesn’t look promising,’” explains Josh Tobin, another OpenAI researcher involved in developing Deep Research. “It’s fascinating to trace those decision paths to grasp how the model is thinking.”
OpenAI clearly views Deep Research as a potential asset for reducing office workloads. “This is something we can scale,” Tobin emphasizes, noting that the agent could be trained to handle specific white-collar tasks. An agent equipped with access to a company’s internal data could efficiently generate a report or presentation, for example. Tobin mentions that the broader objective is to “create an agent not only proficient in compiling reports through web searches but also capable of handling various other tasks.”
Interestingly, since Deep Research was designed to analyze and summarize human-generated text, Tobin mentions that his team was taken aback by the number of users leveraging it to generate code. “It’s a compelling avenue to explore,” he admits. “We’re still figuring out how to interpret this.”
