Рубрики NewsTechnologies

OpenAI introduced Deep Research — a new ChatGPT tool for «deep» online research

Published by Kateryna Danshyna

Deep Research by OpenAI — something like the next agent of artificial intelligence, which specializes in conducting in-depth and comprehensive research using ChatGPT.

In the post, the company notes that the new tool is designed for «people who are engaged in intensive research work» in the fields of finance, politics, or engineering and need «accurate and reliable research». For ordinary users, Deep Research will be useful when making purchases that require «careful study» — such as cars, devices, or even real estate.

An example of Deep Research / OpenAI work

Availability and cost

Deep Research is initially available to owners of ChatGPT Pro for $200 and has a limit of 100 queries; in the future, the tool will be added to Plus, Team, and Enterprise subscriptions.

For now, the tool works exclusively in the web version of ChatGPT, while integration with mobile devices and PCs is promised for this February.

How to use it

To try out a new tool in a ChatGPT chat, select the «deep research» button next to «search» and then enter a query (you can attach files or spreadsheets). The research will take from 5 to 30 minutes, and you will be notified when it is complete.

Currently, Deep Research offers only textual results, but OpenAI intends to add embedded images, data visualization, and other «analytical» results. In addition, it is planned to connect «more specialized data sources».

Learning process and benchmarks

Given that modern AI is still imperfect and can provide misleading results, OpenAI promises that every research result will be «fully documented, with clear citations and a summary to make it easier to verify the information».

Humanity’s Last Exam — an expert-level test consisting of 3000 short-answer and multiple-choice questions on 100 subjects ranging from linguistics to rocketry.

To «enhance» the accuracy of in-depth research, OpenAI uses a special version of models of reasoning o3, «reinforcement-trained» for real-world tasks requiring the use of a browser and Python tools. This type of model building involves learning through trial and error to achieve a specific goal, and when it is achieved — the model receives virtual «rewards» that ideally make it better at the task.

GAIA is a test that evaluates artificial intelligence in real-world questions of three difficulty levels.

OpenAI says that the special version of o3 «used is optimized for web browsing and data analysis» and uses reasoning to find, interpret, and analyze huge amounts of text, images, and PDFs on the Internet. The model is also able to view files uploaded by users, build and repeat graphs using the Python tool, embed generated graphs and images from websites in its answers, and quote specific sentences or passages from sources.