ChatGPT Introduces Deep Research Feature

OpenAI has launched an innovative feature for ChatGPT, named deep research, which enables it to function autonomously. This feature is designed to plan and execute a sequence of steps necessary to gather the required data, adjusting as needed by backtracking and responding to real-time information.

Unlike the previous capabilities that focused merely on generating text, the deep research feature provides a summary of its methodology in a sidebar. This includes citations and a process overview for easier reference.

Users can interact with ChatGPT through various formats: they can ask questions using text, images, and upload additional files like PDFs or spreadsheets for enhanced context. Once a query is made, it usually takes 5 to 30 minutes for the AI to formulate a response, which will eventually feature integrated images and charts in the future.

OpenAI has highlighted some limitations of the deep research tool. For example, it may hallucinate facts, sometimes confounding reliable information with unverified rumors. Additionally, it can struggle with assessing the certainty of its responses.

The push to enhance generative AI tools to offer more value has been a key focus for companies like OpenAI, which asserts that deep research can operate at the level of a professional research analyst. In a demonstration video, a user requests information on the changes in the retail industry over the past three years. The response includes bullet points and tables, showcasing the tool's capability to deliver organized and relevant information.

This new feature comes on the heels of OpenAI's recent launch of Operator, another tool capable of using a web browser to complete tasks autonomously. It bears resemblance to a research prototype named Project Mariner, introduced by Google in December, which has not yet been made public. However, OpenAI's deep research feature is now available, specifically tailored for Pro users.

OpenAI's subscription model allows for up to 100 queries each month for Pro users, while Plus, Team, and eventually Enterprise users will have limited access. The company explains that deep research is resource-intensive, requiring more computational power the longer the task takes. They also promise that all paid users will see increased usage limits once a more efficient version is rolled out.

A press release from OpenAI asserts that the underlying model for deep research achieved a notable accuracy milestone on an AI benchmark called “Humanity’s Last Exam.” This benchmark assesses responses to expert-level queries. The deep research model attained an impressive 26.6 percent accuracy with browsing and Python capabilities enabled. This is significantly higher than the 3.3 percent score of GPT-4o, and also surpasses the next closest model, the o3-mini, which received a 13 percent score based solely on text responses.

AI, Research, Features