OpenAI Unveils Deep Research: A New AI Agent for Extensive Research

OpenAI has introduced a new AI agent named Deep Research, designed specifically for conducting large-scale research tasks. This latest addition utilizes the recently released model, o3, which has been available in a mini version for just two days. With the expanding array of models, agents, and subscription services from OpenAI, it can be challenging to keep track of each one's specific purpose.

At the core of OpenAI's offerings remains ChatGPT. Deep Research functions as an agent integrated within this chatbot framework. According to OpenAI's blog, this agent is capable of performing 'multilayered research on the internet and managing complex tasks' efficiently. It is reported that Deep Research can complete tasks in mere minutes that would typically require several hours for a human to accomplish. However, it is essential to note that this duration does not include the human review time necessary to validate the AI's research results. Like other AI providers, OpenAI emphasizes the importance of having a 'Human-in-the-Loop' to oversee the final output. The company has acknowledged that Deep Research can produce inaccurate information, commonly referred to as 'hallucinations', although they claim that this occurs less frequently compared to other models.

Deep Research is built on a version of the o3 model, which is optimized for web browsing and data analysis. In their blog post, OpenAI highlighted that this development represents a significant advancement toward creating Artificial General Intelligence (AGI), which they believe will accelerate scientific research.

Users can access Deep Research through the standard input field in the web version of ChatGPT, though this feature is limited to individuals with a paid subscription. OpenAI provided an example of how users can leverage this new AI agent to create comparisons of streaming services, a task not typically associated with academic research. The estimated time for answering such queries ranges from 5 to 30 minutes, likely indicating a high operational cost for OpenAI.

Currently, the output produced by Deep Research is text-based, with plans to include images and graphics in the future. OpenAI states that this agent is particularly well-suited for lengthy research projects where accuracy and proper citation are crucial. In contrast, the GPT-4o model is more appropriate for real-time multimodal conversations. Additionally, the new AI agent has demonstrated the ability to answer 26.6 percent of the questions in the benchmark assessment known as Humanity's Last Exam, which focuses on scientific subjects. Previous models reportedly achieved a maximum of only 10 percent, with GPT-4o scoring 3.3 percent and o3-mini-medium and o3-mini-high achieving 10.5 percent and 13 percent, respectively.

Another recently introduced AI agent from OpenAI, known as Operator, also aims to facilitate internet searches and can carry out tasks such as placing orders when provided with credit card information.