ChatGPT has a new feature for conducting in-depth research: what it will change
Kyiv • UNN
OpenAI has introduced the “deep research” feature for ChatGPT, which allows performing complex analytical tasks. The feature is available to Pro users with a limit of 100 queries and demonstrates an accuracy of 26.6% in the “Last Exam of Humanity” test.
OpenAI has introduced a new “deep research” feature for ChatGPT that allows you to perform complex analytical tasks, plan multi-step research, and update answers in real time.
Writes UNN with reference to The Verge.
OpenAI has introduced a new feature for ChatGPT called “deep research” that allows the agent to independently perform complex queries, plan multi-step research, and update answers in real time based on new data.
Unlike a standard text generator, this feature displays a summary of the research process in the sidebar, including sources and key findings.
Users can use not only text queries, but also attach images, PDF files, or spreadsheets to expand the context. Answers are generated within 5-30 minutes, and there are plans to add visualizations such as images and diagrams in the future.
At the same time, OpenAI warns that the function may make mistakes and generate unreliable facts.
Technologies of 2025: how AI and cybersecurity will affect the world01.01.25, 00:45 • 25525 views
Expanding the capabilities of generative AI is one of the key areas of OpenAI. The “deep research” function is positioned as a tool capable of working at the level of a scientific analyst.
For example, in a demo video, the agent analyzes changes in the retail industry over the past three years and provides a structured response with tables and bullet points. The feature follows the launch of OpenAI Operator, which uses a web browser to perform tasks and is reminiscent of Google's research project called Project Mariner. The latter has not yet been made available to the public, while OpenAI has already started testing deep research in the Pro version.
Currently, access is limited: subscribers to the $200 per month plan receive 100 requests, and Plus, Team, and Enterprise users are promised expanded access in the future.
According to OpenAI, the model used for in-depth research has reached a new level of accuracy in the “Last Exam of Humanity” test.
Using a browser and Python, it demonstrated an accuracy of 26.6%, which is significantly higher than GPT-4o (3.3%) and the second most accurate model, o3-mini (13%). OpenAI plans to further improve the efficiency and accessibility of this technology.
Recall
OpenAI CEO Sam Altman acknowledged that the Chinese chatbot DeepSeek has weakened OpenAI's leadership in the AI industry and said that he believes OpenAI is “on the wrong side of history” in terms of open-sourcing its technology.
