OpenAI has unveiled a new AI-powered tool, “deep research,” designed to enhance complex, in-depth research capabilities within its ChatGPT platform. The feature targets professionals in fields such as finance, science, policy, and engineering, as well as consumers conducting high-stakes research for purchases like cars or appliances.
In a blog post, the company stated the tool is tailored for scenarios requiring “thorough, precise, and reliable research” beyond quick summaries, enabling users to “assiduously consider information from multiple websites and other sources.”
Availability and Access
The feature is now available to ChatGPT Pro users, capped at 100 queries monthly, with plans to expand access to Plus and Team subscribers “in about a month,” followed by Enterprise clients. OpenAI noted query limits for paid tiers will soon be “significantly higher.”
However, users in the U.K., Switzerland, and the European Economic Area face an undetermined wait, as regional rollout timelines remain unspecified.
To use deep research, users select the “deep research” option in ChatGPT’s composer, input a query, and optionally attach files or spreadsheets. Responses may take 5 to 30 minutes, with notifications upon completion. Currently, web-only, mobile, and desktop integration is slated for late July.
Features and Roadmap
While initial outputs are text-based, OpenAI plans to introduce embedded images, data visualizations, and analytical elements soon. Future updates will also enable connections to “more specialized data sources,” including subscription-based and internal resources.
Accuracy and Model Architecture
Addressing concerns about AI reliability, OpenAI emphasized that every deep research output will be “fully documented, with clear citations and a summary of [the] thinking, making it easy to reference and verify the information.” The tool leverages a specialized version of the o3 reasoning model, trained via reinforcement learning on tasks requiring web browsing and Python-based data analysis.
This approach allows the model to “pivot as needed in reaction to information it encounters,” browse user-uploaded files, generate graphs, and cite specific sources.
In performance testing, the o3 model achieved 26.6% accuracy on Humanity’s Last Exam, a benchmark with over 3,000 expert-level questions. While this score appears low, OpenAI notes the exam is intentionally rigorous, surpassing benchmarks like Gemini Thinking (6.2%), Grok-2 (3.8%), and GPT-4o (3.3%).
Limitations and Ethical Considerations
OpenAI acknowledges the tool’s imperfections, including potential errors, difficulty distinguishing authoritative sources from rumors, and occasional formatting issues in citations. The company also flagged its tendency to “fail to convey uncertainty.”
For educators and information integrity advocates, deep research’s cited outputs may offer a preferable alternative to uncrafted chatbot summaries. However, OpenAI’s transparency measures hinge on users rigorously verifying results, a step that remains uncertain in practice. As the company stated, the broader impact depends on whether users treat the tool as a “professional-looking text to copy-paste” or engage in genuine analysis.
