OpenAI has launched a new offering known as “Deep Research,” which can plan and execute multi-step online research tasks. Potential uses of the product, as OpenAI outlined in an announcement include navigating scientific literature or analyzing UI/UX research data. In an internal evaluation, the agent shows strength in automating routine online research tasks, achieving nearly 20% pass rates on low-complexity queries while struggling with more economically valuable challenges.
[OpenAI]
Early feedback on social media is mixed with some critics calling the research quality “mediocre” and noted a lack of citations in one sample. OpenAI CEO Sam Altman was optimistic about its potential, calling it “a huge step forward for AI,” and saying that a more powerful version was in the works.
[Deep Research can do] a single-digit percentage of all economically valuable tasks in the world…. You can now have an army of research assistants at your disposal to do anything you like. And we are going to take this much further.
On Humanity’s last exam, a test spanning more than 3,000 expert-level questions across more than 100 disciplines, Deep Research achieved 26.6%, more than double that of the DeepSeek R1 model’s 9.4% score.
Deep Research’s architecture enables it to break down complex research queries into manageable chunks. The system can process both web-based resources and uploaded documents. It taps a browser- and Python-tool layer to sift through public websites, PDFs, and images, then adjust its research plan as needed. This architecture, built on a version of OpenAI’s upcoming o3 model optimized for search and synthesis, can potentially turn hours of manual digging into short, detailed reports. It also provides citations for its research.
“Deep research is built for people who do intensive knowledge work in areas like finance, science, policy, and engineering and need thorough, precise, and reliable research,” OpenAI noted when debuting the feature.
Currently, access to Deep Research is reserved for ChatGPT Pro users, and limited to 100 queries per month, though OpenAI plans on rolling out access to other paid accounts in the coming months.
OpenAI notes, however, that the tool does have shortcomings. Most notably, it can still “hallucinate,” or make up facts when struggling with context. For instance, it might conflate authoritative sources with unverified information, making manual verification of data still a must.
