OpenAI: New ChatGPT Agent For Deep Research

As China’s DeepSeek chatbot intensifies rivalry in the artificial intelligence space, US tech firm OpenAI on Monday revealed a ChatGPT tool termed “deep research” that can provide comprehensive reports. A new AI “agent” from OpenAI is being unveiled to assist users in carrying out intricate, in-depth research using ChatGPT, the company’s AI-powered chatbot platform. It’s appropriately named “deep research”.

The statement was made in Tokyo, where Sam Altman, the CEO of OpenAI, also announced a new partnership with tech investor SoftBank Group to provide businesses with cutting-edge AI services.

Silicon Valley is in a frenzy about AI newcomer DeepSeek, which some have dubbed a wake-up call for US developers because of its strong performance and alleged low cost.

This new capacity was created for “people who do intensive knowledge work in areas like finance, science, policy, and engineering and need thorough, precise, and reliable research,” according to a blog post published by OpenAI on Sunday. Anybody making “purchases that typically require careful research, like cars, appliances, and furniture,” the business noted, would find it helpful.

According to OpenAI, whose ChatGPT spearheaded the public’s awareness of generative AI in 2022, their latest tool “completes in tens of minutes what would take a human many hours.”

In essence, ChatGPT deep research is meant for situations in which you need to carefully evaluate data from several websites and other sources rather than relying just on a summary or fast response.

According to OpenAI, ChatGPT Pro customers will now have access to deep research, up to 100 questions each month. Plus and Team users will follow, then Enterprise. (The business stated that the query limitations for paying customers should be “significantly higher” in the near future, and that OpenAI plans to roll out Plus in around a month.) The launch is geo-targeted; OpenAI was unable to provide a release schedule for ChatGPT users in the European Economic Area, Switzerland, and the United Kingdom.

The business also stated in a statement that “you give it a prompt, and ChatGPT will find, analyse, and synthesise hundreds of online sources to create a comprehensive report at the level of a research analyst.”

Simply choose “deep research” in the composer and type in a query, with the ability to attach files or spreadsheets, to utilize ChatGPT deep research. (At this time, it is only available online; later this month, PC and mobile app connectivity will be added.) You will receive a message when the search is finished, and it may take five to thirty minutes to find the answer to the query.

The results of ChatGPT deep research are currently limited to text. However, OpenAI stated that it plans to soon provide data visualizations, embedded pictures, and other “analytic” outputs. Connecting “more specialized data sources,” such as “subscription-based” and internal resources, is another item on the roadmap, according to OpenAI.

The main query is: How accurate is ChatGPT’s in-depth analysis? After all, AI is not flawless. It is prone to mistakes such as hallucinations, which might be especially detrimental in a situation involving “deep research.” Perhaps for this reason, OpenAI said that all ChatGPT deep research outputs will be “completely documented, with clear citations and a summary of [the] thinking, making it easy to reference and verify the information.”

Whether those mitigations will be enough to counteract AI errors is still up in the air. ChatGPT Search, OpenAI’s AI-powered online search function, regularly makes mistakes and provides incorrect responses to queries. According to TechCrunch’s testing, for some questions, ChatGPT Search yielded fewer helpful results than Google Search.

OpenAI is employing a customized version of its newly revealed o3 “reasoning” AI model, which was trained via reinforcement learning on “real-world tasks requiring browser and Python tool use,” to increase the accuracy of deep research. In essence, reinforcement learning “teaches” a model to accomplish a certain objective through trial and error. The model receives virtual “rewards” as it approaches the objective, which should improve its performance in the future.

This version of o3 is “optimized for web browsing and data analysis,” according to OpenAI, which also stated that it uses reasoning to search, understand, and analyze vast volumes of text, photos, and PDFs on the internet, changing course as necessary in response to new information.” According to the business, the model can also “plot and iterate on graphs using [a Python] tool, embed both generated graphs and images from websites in its responses, and cite specific sentences or passages from its sources.” It can even “browse over user-uploaded files.”

According to OpenAI, it used Humanity’s Last Exam, an assessment consisting of over 3,000 expert-level questions covering a range of academic subjects, to evaluate ChatGPT’s deep research. The accuracy of the o3 model that drives deep research was 26.6%, which may seem like a failing grade, but Humanity’s Last Exam was intended to be more difficult than other benchmarks in order to keep up with model improvements. The deep research o3 model outperformed Grok-2 (3.8%), OpenAI’s own GPT-4o (3.3%), and Gemini Thinking (6.2%), according to OpenAI.

However, OpenAI points out that ChatGPT’s deep research has limits, occasionally producing inaccurate conclusions. According to the firm, deep research can make formatting mistakes in reports and citations, struggle to discern reliable facts from hearsay, and frequently fail to communicate when it is unsure about something.

This kind of thorough, well referenced output likely sounds more enticing than a misleadingly straightforward chatbot summary devoid of citations to anyone concerned about the effects of generative AI on students or anyone looking for information online. We’ll see, though, if the majority of users will truly examine and double-check the output or if they just copy and paste it as a more polished text. If all of this seems familiar, Google just unveiled a comparable AI capability less than two months ago with the same moniker.