OpenAI has announced deep research, a new artificial intelligence agent designed to help people conduct «deep, complex research» using ChatGPT.
OpenAI has announced deep research, a new artificial intelligence agent designed to help people conduct «deep, complex research» using ChatGPT.
In a blog post, OpenAI said the new feature is aimed at «people who work intensively with knowledge in fields like finance, science, politics, and engineering and need thorough, accurate, and reliable research.» The company also added that it could be useful for those making «purchases that typically require thorough research, such as cars, appliances, and furniture.»
Basically, ChatGPT’s deep research is intended for cases where the user doesn’t just want to get a quick answer or summary, but rather carefully consider information from multiple websites and other sources, TechCrunch notes.
OpenAI said it is making deep research available to ChatGPT Pro users today, with a limit of 100 queries per month, with support for Plus and Team users, and Enterprise later. The company plans to roll out Plus in about a month, and query limits for paid users should be «significantly higher» soon. The new AI agent is currently unavailable to ChatGPT users in the UK, Switzerland, and the European Economic Area.
To use ChatGPT for deep research, you need to select deep research and then enter a query with the ability to attach files or spreadsheets. This feature is currently only available on the web version, with integration with mobile and desktop apps coming later this month. Deep research can take anywhere from 5 to 30 minutes to answer a question — the user will receive a notification when the search is complete.
For now, ChatGPT’s deep research results are text-only. But OpenAI says it plans to add embedded images, data visualizations, and other «analytical» results soon. The roadmap also includes the ability to connect to «more specialized data sources,» including «subscription-based» and internal resources.
OpenAI added that each result of ChatGPT’s deep research will be «fully documented, with clear citations and a concise statement of the thought, allowing for easy referencing and verification.»
To improve the accuracy of deep learning, OpenAI is using a special version of its recently announced o3 «reasoning» AI model that has been trained using reinforcement learning on «real-world tasks that require the use of a browser and Python tools.» Reinforcement learning essentially «trains» a model through trial and error to achieve a specific goal. As the model gets closer to the goal, it receives virtual «rewards,» which ideally help it perform better at future tasks.
OpenAI said it tested ChatGPT’s deep research using the «Last Exam of Humanity,» a test that includes more than 3,000 expert-level questions from a variety of academic disciplines. The o3 model on which deep research is based achieved an accuracy of 26.6 percent, which may seem like a poor score, but the «Last Exam of Humanity» was designed to be more rigorous than other benchmarks to keep pace with model improvements. According to OpenAI, the o3 deep research model outperformed Gemini Thinking (6.2 percent), Grok-2 (3.8 percent), and its own GPT-4o (3.3 percent).
However, OpenAI notes that ChatGPT’s deep research has limitations, sometimes making mistakes and drawing incorrect conclusions. The company says that deep research can struggle to distinguish between true information and rumors, and often doesn’t make it clear when the AI is unsure, and can also make formatting errors in reports and citations.
Recall that in late January, Sam Altman, CEO of the American company OpenAI, commented on the hype surrounding artificial intelligence models from the Chinese DeepSeek. He called them «amazing,» but assured that his company can do better.
Microsoft later made OpenAI’s o1 model free for all Copilot users. The model powers the Think Deeper feature, allowing Copilot to analyze complex questions and generate deeper answers.
