OpenAI has announced a new artificial intelligence tool specializing in deep research, named “Deep Search.” According to Reuters, the company stated that this tool conducts multi-step searches on the internet to accomplish complex tasks.
The company explained that the new deep search tool is powered by a version of the “OpenAI o3” model, which is specifically designed for web browsing and data analysis. OpenAI claims that when a user submits a query in ChatGPT, the chatbot will search, analyze, and summarize multiple sources, including texts, images, and PDF files, to generate a comprehensive report that adheres to research standards.
In a blog post, OpenAI stated that this new tool is designed for individuals conducting intensive research and studies in various fields such as science, engineering, politics, and finance, requiring precise, comprehensive, and reliable research. Additionally, it can be beneficial for anyone making purchase decisions that demand thorough research, such as cars, household appliances, and furniture.
Availability and Usage
The “Deep Search” tool will first be available to Pro users with a maximum limit of 100 queries per month. It will then be rolled out to Plus and Team users within approximately a month, followed by Enterprise users.
To use this tool, users simply need to select “Deep Search” in the message box and enter their query. There is also an option to attach files or spreadsheets. The deep search process can take anywhere from 5 to 30 minutes to complete, and users will receive a notification once the search is finished. A sidebar will then display a summary of the steps taken and the sources used.
It is worth noting that the tool is currently limited to web use only, with plans to release it for mobile and desktop applications later this month.
How Accurate Is “Deep Search”?
OpenAI claims that its new tool can complete tasks in 10 minutes that would take an average person several hours. However, the tool is still in its early stages and naturally has limitations and challenges. It may struggle to distinguish between reliable information and rumors and currently exhibits weaknesses in confidence calibration, often encountering difficulties in identifying unreliable data.
Despite these challenges, OpenAI assures that all “Deep Search” outputs in ChatGPT will be fully documented with clear sources, ensuring transparency and verifiability. However, the company has not specified whether these measures are sufficient to combat AI-related errors. AI-powered web search tools frequently make mistakes and provide inaccurate answers. According to TechCrunch, a test found that ChatGPT Search delivered less useful results compared to Google Search.
To enhance the accuracy of deep searches, OpenAI utilizes a specialized version of its reasoning AI model, “o3,” trained through reinforcement learning on real-world tasks requiring browser usage and Python programming.
The company also revealed that it tested “Deep Search” using the “Humanity’s Last Exam” model, a rigorous evaluation containing over 3,000 expert-level questions across a wide range of academic fields. The “o3” model, which powers deep search, achieved an accuracy rate of 26.6%. While this may seem low, “Humanity’s Last Exam” is designed to be more challenging than other benchmarks to drive advancements in AI models.
According to OpenAI, this score is significantly higher than other AI models, such as “Gemini Thinking” at 6.2%, “Grok-2” at 3.8%, and “GPT-4o” at 3.3%.
Although AI models still suffer from inaccuracies, for those concerned about the impact of generative AI on students or anyone seeking reliable information online, this kind of deep and source-backed output may be more appealing than a simple, citation-free summary. However, it remains to be seen whether users will critically analyze and review the output or simply treat it as a professional text ready to be copied and pasted.
