OpenAI Launches ChatGPT Agent for In-Depth Research

OpenAI Introduces a New AI Agent for Comprehensive Research

OpenAI has announced the release of a novel AI agent, designed to facilitate detailed and complex investigations utilizing its ChatGPT platform. This new capability aims to assist users in conducting thorough research.

Introducing "Deep Research"

The newly launched feature, aptly named “deep research,” is intended for professionals engaged in intensive knowledge work. These fields include finance, science, policy, and engineering, where precise and dependable research is paramount.

It can also prove beneficial for consumers making significant purchases, such as automobiles, appliances, or furniture, which typically necessitate careful consideration.

Functionality and Intended Use

ChatGPT deep research is specifically designed for scenarios demanding more than just a quick response or summary. It allows users to meticulously evaluate information gathered from numerous websites and diverse sources.

Currently, the feature is available to ChatGPT Pro subscribers, initially limited to 100 queries monthly. Expansion to Plus and Team users is planned for the coming weeks, followed by Enterprise access.

OpenAI anticipates a Plus rollout within approximately one month, with increased query allowances for paying users to be implemented shortly thereafter. The initial launch is geographically targeted, excluding customers in the U.K., Switzerland, and the European Economic Area for the time being.

To initiate a deep research session, users simply select the “deep research” option within the composer and input their query. Files and spreadsheets can also be attached. The process, currently web-based, is expected to integrate with mobile and desktop applications later this month.

Output and Future Enhancements

The time required for ChatGPT deep research to complete a query ranges from five to thirty minutes, with users receiving a notification upon completion.

Presently, outputs are text-based. However, OpenAI intends to incorporate embedded images, data visualizations, and other analytical outputs in the near future. The roadmap also includes connectivity to specialized data sources, including subscription-based services and internal resources.

Accuracy and Reliability

A critical consideration is the accuracy of ChatGPT deep research. Recognizing the inherent imperfections of AI, including the potential for hallucinations and errors, OpenAI emphasizes full documentation of all outputs.

Each response will include clear citations and a summary of the reasoning process, enabling easy verification of the information provided.

Previous testing of OpenAI’s AI-powered web search within ChatGPT, known as ChatGPT Search, has revealed instances of inaccuracies and incorrect answers. Comparisons with Google Search have shown varying levels of usefulness depending on the query.

Underlying Technology

To enhance accuracy, OpenAI is leveraging a specialized version of its recently unveiled o3 “reasoning” AI model. This model was trained using reinforcement learning on real-world tasks involving browser and Python tool utilization.

Reinforcement learning employs a trial-and-error approach to achieve specific goals, with virtual “rewards” guiding the model towards improved performance. This version of o3 is specifically optimized for web browsing and data analysis.

It utilizes reasoning capabilities to search, interpret, and analyze extensive amounts of text, images, and PDFs online, adapting its approach based on encountered information. The model can also process user-uploaded files, generate and embed graphs, and cite specific passages from sources.

Performance Evaluation

OpenAI evaluated ChatGPT deep research using Humanity’s Last Exam, a challenging assessment comprising over 3,000 expert-level questions across various academic disciplines.

The o3 model achieved an accuracy of 26.6%. While seemingly low, the exam is intentionally designed to be more rigorous than standard benchmarks. The model outperformed Gemini Thinking (6.2%), Grok-2 (3.8%), and OpenAI’s GPT-4o (3.3%).

Limitations and Considerations

Despite these advancements, OpenAI acknowledges limitations, including potential errors and incorrect inferences. The system may struggle to differentiate between credible information and rumors, and may not always indicate uncertainty. Formatting errors in reports and citations are also possible.

This type of in-depth, well-cited output may be particularly appealing to those concerned about the impact of generative AI on education and information access. However, the extent to which users will critically analyze and verify the output remains to be seen.

Notably, Google announced a similar AI feature with the same name less than two months prior.