OpenAI's ‘Deep Research’ Can Actually Make Professional Reports With Citations

Not to be out-done by Deepseek, OpenAI is launching a new Deep Research feature in ChatGPT. This is OpenAI's newest Agentic AI feature (after Operator), which builds on the recent trend of making AI more autonomous. According to OpenAI, Deep Research is capable of producing detailed reports matching the level of a research analyst. In a layperson's terms, it browses and interprets the internet for you.

Deep Research uses OpenAI's upcoming o3 reasoning model to perform complex tasks, taking its own sweet time to do so. The feature is available now for ChatGPT Pro customers (the pricey sub that costs $200/month), but will soon be available for ChatGPT Plus and Enterprise users as well.

How OpenAI's Deep Research AI Agent works

OpenAI's Deep Research tool is designed to work independently from you. You give it a detailed prompt, after which it'll ask some clarifying questions. Then, it will go and do its own thing in the background. According to OpenAI, a Deep Research stint can last anywhere between 5 minutes and 30 minutes, but the company claims it's able to do multiple hours worth of human-level work in the span within just a dozen or so minutes.

While it's working, there's a panel on the right side of the page that shows everything it's doing, live. Think of this as the bot's citations, but it also explains its "thought process." It can connect to the internet, search online, read web pages, and analyze or synthesize massive amounts of information in the form of text, images, and PDFs. All of this is a bit compute-intensive, so OpenAI is limiting Pro users to just 100 queries a month. A smaller, more efficient model will be rolled out in the coming months, as well.

The Deep Research feature is purpose-built for knowledge workers in the field of science, finance, engineering, and policy. But OpenAI says that it can be equally useful for consumers too. OpenAI gave an example of how Deep Research can help perform hyper-personalized research for big shopping decisions. Things like helping you decide between cars, furniture, appliances, or electronics. Since the tool can synthesize information from thousands of articles and reviews, it can supposedly build a report customized to your needs.

Credit: OpenAI

According to OpenAI, "deep research was rated by domain experts to have automated multiple hours of difficult, manual investigation"

OpenAI offers multiple examples where Deep Research's insights can be valuable to users, saving hours of research time. The company says it can be used to understand extremely niche and specific problems via scientific studies and journals.

Credit: OpenAI

For example, a Chemistry prompt asks ChatGPT to "discuss the differences between pure- and mixed-gas sorption for glassy polymers, how the dual-mode sorption model can be used to predict mixed-gas sorption behavior in glassy polymers," the model then goes on to understand sorption models, accesses open-source information, clarifies key problems, pulls up PDFs, and even refines the model before piecing together all the content. According to OpenAI, this task helped save 4 hours of time.

OpenAI's post also highlights similar use cases for Deep Research in the healthcare industry and linguistics, saving five hours and two hours, respectively.

Deep Research also supposedly performed well on Humanity's Last Exam, an AI benchmark, testing expert-level knowledge across more than 100 fields. Deep Research scored 26.6% accuracy, the highest score yet on the text. By comparison, DeepSeek-R-1 scored 9.4%, and GPT-4o managed just 3.3%.

While Deep Research is based on a reasoning model, and not an LLM, it still uses a language model to work with the input, and generate the output text. OpenAI warns that the Deep Research model can still hallucinate and make up facts, so it's still better to keep an eye on the research output, and not to trust it blindly.

View the full article