OpenAI is taking AI-powered research to a new level with the launch of Deep Research, a feature within ChatGPT that can conduct multi-step online investigations and compile comprehensive reports in minutes—tasks that might take human hours or even days.

Deep Research is designed to independently explore, analyze, and synthesize information from across the web, behaving more like a research assistant than a simple query responder. Whether it’s conducting a competitive analysis of the streaming industry or investigating the best commuter bike based on niche user preferences, OpenAI’s claims the new tool aims to dig deeper than a basic search engine.

Unlike traditional AI models that generate quick summaries based on pre-trained data, Deep Research actively browses the web, interpreting live information, cross-referencing sources, and even analyzing uploaded documents. It doesn’t just pull data—it reasons through conflicting information, adjusts its approach as needed, and cites sources for transparency.

This is all powered by a specialized version of OpenAI’s upcoming o3 model, optimized for web browsing and complex data analysis. The model leverages reinforcement learning techniques to plan its research path, ensuring it finds the most relevant information rather than simply regurgitating surface-level results.

Users can activate Deep Research within ChatGPT by selecting the feature in the message composer. Once a query is entered, a sidebar will track the AI’s progress, outlining the steps taken and listing its sources. Depending on the complexity of the request, the research process may take anywhere from 5 to 30 minutes.

OpenAI sees this as more than just a convenience tool. In its announcement, the company framed Deep Research as part of a longer-term vision: the development of artificial general intelligence (AGI) capable of producing novel scientific discoveries rather than just summarizing existing knowledge.

“The ability to synthesize knowledge is a prerequisite for creating new knowledge,” OpenAI stated in its launch post, hinting at a future where AI might not just compile research—but contribute to it.

OpenAI claims Deep Research significantly outperforms existing AI models on complex research tasks. On "Humanity’s Last Exam," a benchmark that tests AI across 3,000 expert-level questions in subjects ranging from linguistics to rocket science, Deep Research scored 26.6% accuracy—a major leap over OpenAI’s previous model (o1, which scored 9.1%) and well ahead of competitors like GPT-4o (3.3%) and Claude 3.5 Sonnet (4.3%).

It also topped the leaderboard on GAIA, an external benchmark evaluating AI’s ability to answer real-world questions that require reasoning and web research.

Despite its strengths, OpenAI admits Deep Research isn’t perfect. The model can hallucinate information (though at a lower rate than previous versions), struggle with misinformation, and sometimes fail to communicate uncertainty accurately. Early users may also encounter formatting errors in reports and citations.

Additionally, access is currently limited. At launch, only Pro users of ChatGPT can access Deep Research, with Plus and Team users expected to follow soon. The UK, Switzerland, and the European Economic Area won’t have access at first due to regulatory constraints.