ChatGPT agent can now perform deep research for you.

Image: OpenAI

Images: Users can add context to their questions by adding images, text, or PDFs. It will then take between 5 and 30 minutes to create a response in the chat window. In the future, it will be able embed images and charts. OpenAI also points out limitations of deep research. It can “sometimes hallucinate”and make up facts. It has difficulty telling the difference between authoritative information and rumors.

Companies like OpenAI have promised agents that they will develop ways to make generative AI tools more useful and worthwhile. They also claim that deep research can operate at the level as a research analyst. The demo video begins with an inquiry about changes in the retail sector over the past three years. The response includes bullet points and table.

The feature is similar to Project Mariner, the research prototype Google demonstrated in December. It closely follows OpenAI’s launch of Operator, which uses a web-browser to perform tasks for you. Google’s tool has not been made public yet, but Deep Research is now available “with a version that is optimized for Pro users.”

OpenAI offers up to 100 queries a month to those who pay the $200 monthly fee, with “limited access” promised to Plus, Team and Enterprise users. They call the ability “very computation intensive,” as it requires more inference computation the longer the research takes. The company also states that paid users will receive higher rate limits when a faster and more cost-effective version becomes available.

According to a press release, the model that powers deep research has achieved a new high in accuracy on an AI benchmark dubbed “Humanity’s Last Exam ”, which asks expert-level questions. OpenAI’s deep research model achieved an accuracy of 26.6% with browsing and python enabled, which is well above GPT-4o’s 3.3 %. The next highest scorer was its o3 mini (high) model, evaluated only on text.