OpenAI has unveiled its latest AI research assistant, ‘Deep Research’, offering a preview of how future AI agents may operate. Powered by OpenAI’s o3-mini model, which has been trained to use trial and error to tackle complex queries, Deep Research marks one of OpenAI’s first significant steps towards creating autonomous AI capable of following instructions and conducting independent research, says Mark Sullivan, a senior writer at Fast Company, covering emerging technology, AI, and tech policy.
Designed for professionals in finance, science, policy, and engineering, Deep Research delivers thorough, precise, and reliable information. It can also assist with major purchase decisions, such as buying a house or car. However, due to its high computational demands, OpenAI has restricted access to Pro users on its $200-per-month plan, limiting them to 100 searches per month. To provide a hands-on experience, OpenAI granted me access to the tool for a week, where I discovered a new ‘Deep Research’ button beneath the standard ChatGPT prompt window.
To test its capabilities, I first asked Deep Research to investigate non-drug products that claim to relieve low back pain. While I had consumer technology in mind, I did not specify this, leading ChatGPT to clarify whether I wanted to include ergonomic furniture and posture correctors. Within six minutes, the model analysed 20 sources and generated a 2,000-word report, detailing consumer devices such as heated vibration belts, contact pad systems, and Transcutaneous Electrical Nerve Stimulation (TENS) units. Impressively, it even produced a comparison grid, outlining key specifications and pricing for ten different devices. Lacking expertise in this area, I could not identify any missing information or inaccuracies in the results.
To push the model further, I posed a more complex query:
“I would like an executive overview of the latest research on using artificial intelligence to discover new cancer treatments or diagnostic tools. Please emphasise the most promising treatments that are closest to clinical application.”
Similar to DeepSeek’s R1 model and Google’s Gemini Advanced 2.0 Flash Thinking Experimental, OpenAI’s research tool employs ‘chain-of-thought’ reasoning, allowing users to track how the AI formulates its response. As it processed my request, it displayed a step-by-step breakdown, indicating that it was analysing AI’s role in cancer diagnostics and treatment, covering medical imaging, pathology, genomics, and radiotherapy planning. OpenAI made a smart UX decision by placing this reasoning process in a side panel, rather than overlaying it on the research results. However, a notable drawback is that the chain-of-thought process disappears once the research is complete, providing only a fleeting glimpse into how the AI arrives at its conclusions.
