What’s better than an AI chatbot that can assist you with tasks? One that can do them for you. OpenAI continues to build out its AI agents in ChatGPT with the launch of Deep Research.
Deep Research
On Sunday, OpenAI unveiled Deep Research, an AI agent that can conduct multi-step research for you by pulling a robust amount of information from the web and synthesizing those sources for you in a comprehensive report. Once prompted, Deep Research can work entirely independently; it’s like having a research analyst at your command.
Today, we launch our next agent that can do work for you independently — deep research. Give ChatGPT a prompt, and it will find, analyse & synthesize hundreds online sources to create an extensive report in tens minutes instead of what would take a person many hours. pic.twitter.com/03PPi4cdqi
– OPENAI (@OpenAI)””https://twitter.com/OpenAI/status/1886219085236850889?ref_src=twsrc^tfw””> Powering Deep Research, a version optimized for web browsing and analysis of data, is an OpenAI o3 model that will be available on February 3, 2025. By leveraging the advanced reasoning capabilities of o3, it can search, interpret, and output massive amounts content from the internet, including images, texts, and more.
Depending on the task, each report can be generated in 5 to 30 minutes. During this time, you can do other tasks, optimizing workflow productivity. The report is then outputted in the chat. In the coming weeks, the agent may also include data visualizations and images.
See how Gen AI can improve customer experience – one bank’s method
OpenAI claims that the same task would take humans several hours. The agent is also designed to be good at finding niche data that would require humans performing multiple searches.
OpenAI states that Deep Research is aimed at those who perform intensive knowledge work, such as in finance, science and policy, or engineering, and require thorough, reliable research. Each report includes a summary of what the agent thought and clear citations so that users can verify the information themselves.
It is a good idea to double-check a chatbot’s responses, as they are prone hallucinations. OpenAI warns Deep Research “can sometimes hallucinate facts in responses or make incorrect inferences, though at a notably lower rate than existing ChatGPT models, according to internal evaluations.” OpenAI added that the agent could struggle to distinguish authoritative from rumors, and can fail correctly to convey uncertainty, highlighting the necessity for human review.
Performance comparison
OpenAi’s Blog Post that announced the feature includes the same results side-by-side of GPT-4o and Deep Research in order to demonstrate how the same prompt can produce very different results. The results generated by Deep Research were more robust and organized.
Screenshot by Sabrina Ortiz/ZDNET
Deep Research also outperformed GPT-4o on Humanity’s Last Exam, a recently launched AI benchmark exam by Scale AI and the Center for AI Safety (CAIS) that tests various subjects on expert-level questions. Deep Research scored a 26.6% accuracy, outperforming GPT-4o, Grok-2, Claude 3,5 Sonnet, Gemini Thinking, o1, and even o3-mini high, which had just scored the highest score a couple of days prior, as highlighted by OpenAI CEO Sam Altman.
On friday, “humanity’s last exam” had a high score of 13%. Now on sunday, deep analysis gets 26.6%.
– Alone Altman, @sama””https://twitter.com/sama/status/1886220281565381078?ref_src=twsrc%5Etfw””> OpenAI published the performance results of Deep Research on February 3, 2025. This included GAIA, an external benchmark that evaluates AI in real-world situations, and an internal evaluation of expert level tasks across different areas within deep research. Deep Research achieved impressive results in both evaluations, topping the GAIA leaderboard.
Access
Due to the computing power needed to run Deep Research, only ChatGPT users can access it for the time being. The $200 per month subscription includes up to 100 queries in an optimized version, as well as other benefits such unlimited access to ChatGPT, Sora, and Operator. This AI agent can perform basic browser tasks such reservations. ChatGPT Plus, Team and Enterprise users will be the first to get access. OpenAI plans to release a more cost-effective, faster version of the feature based on a smaller but equally efficient model.
See how Gen AI can improve customer experience – one bank’s example
Google offers a similar feature called Deep Research that is free to all its Gemini Advanced users. Google One AI Premium costs $20 per month. Altman replied to a question in December. The X user asked Altman to “do a deep research feature like Gemini but better,” and “kk,” suggesting the newly released Deep Research is OpenAI’s response to Google.
Microsoft announced last week a feature called Think Deeper that allows users to leverage OpenAI’s O1 reasoning to deliver better-quality answers to complex prompts. It does not have internet access or agentic capabilities, as with OpenAI’s Deep Research and Gemini’s Deep Research. The best part is that it’s completely free.
