This week, OpenAI has rolled out its latest innovation—Deep Research—an advanced agent integrated into ChatGPT that promises to revolutionize how professionals and consumers alike tackle complex, multi-step research tasks. Unlike conventional models that deliver quick responses, Deep Research is designed to take its time—ranging from 5 to 30 minutes—to produce fully cited, comprehensive reports that mirror the work of expert analysts.
Deep Research represents a significant leap forward in autonomous, agentic AI. Built on a refined version of OpenAI’s upcoming o3 model, the tool is engineered to comb through the internet, synthesizing information from hundreds of sources to deliver detailed reports on a wide array of topics. Whether it’s performing a market analysis on mobile adoption trends, comparing niche products, or delving into academic research, the system methodically plans its approach, asking clarifying questions along the way to ensure precision.
In a demonstration held in Tokyo, team members showcased how the tool could independently generate a research report complete with tables, graphs, and extensive citations. For instance, one query tasked Deep Research with analyzing global mobile penetration rates and language learning trends, a challenge that traditionally requires hours of manual investigation.
Deep Research’s capabilities are built on end-to-end reinforcement learning methods honed on demanding browsing and reasoning tasks. Key features include:
The tool’s ability to ask targeted clarifying questions ensures that even open-ended or complex queries are broken down into actionable research steps, much like a human expert would do.
In rigorous internal evaluations, Deep Research has already shown promising results. Notably, on “Humanity’s Last Exam”—a benchmark featuring over 3,000 expert-level questions across more than 100 subjects—the model achieved an impressive accuracy of 26.6%, far surpassing earlier iterations. It also set new records on the GAIA benchmark, which measures AI performance on real-world, multi-modal tasks.
These benchmarks indicate that the model excels not only in technical domains like coding and math but also in areas that require nuanced reasoning and extensive context, such as humanities and social sciences.
Despite its breakthrough capabilities, Deep Research is not without challenges. Early users may encounter occasional factual hallucinations or minor formatting issues in the generated reports. Additionally, the model can sometimes struggle with conveying uncertainty, especially when distinguishing between well-established facts and less reliable information. Given its compute-intensive nature, tasks can also take a considerable amount of time to complete.
However, OpenAI acknowledges these limitations and expects iterative improvements as the tool is exposed to more real-world usage.
For now, Deep Research is being offered exclusively to ChatGPT Pro users, who receive up to 100 queries per month. Access for Plus and Team users is slated to follow shortly, with further rollouts planned for education and enterprise sectors. Notably, availability in regions such as the United Kingdom, Switzerland, and the European Economic Area will be introduced in subsequent updates.
Looking ahead, future iterations of Deep Research are expected to feature support for mobile and desktop apps, as well as connectivity with subscription-based or internal data sources. This will further enhance its ability to deliver personalized, robust outputs. There is also a vision for a broader ecosystem of agentic experiences—where Deep Research collaborates seamlessly with other operational tools to execute complex, real-world tasks asynchronously.
OpenAI’s Deep Research is poised to redefine the landscape of knowledge work by automating time-consuming research processes and delivering expert-level reports with unprecedented depth and clarity. As the technology continues to mature, it is likely to become an indispensable asset for professionals in finance, science, policy, engineering, and beyond—transforming how we approach and interact with the vast resources of the internet.
Read OpenAI's announcement here.
