ChatGPT Deep Research: A New Tool for Deep and Autonomous Data Exploration

ChatGPT’s New "Deep Research" Agent: A Powerful Tool for Autonomous Research

OpenAI has recently unveiled a groundbreaking feature for ChatGPT, called deep research, which allows the AI to conduct detailed, multi-step research on your behalf. Unlike previous iterations, which focused on generating text responses based on existing knowledge, deep research operates autonomously, planning and executing a trajectory to find and verify the necessary data. It can even backtrack and adjust its approach based on real-time information, offering a more dynamic and thorough research process.

How It Works

When a user submits a query, whether through text, images, or even additional files like PDFs and spreadsheets, ChatGPT’s deep research agent begins by sourcing relevant data and structuring a research plan. The process is displayed in a sidebar, offering transparency on how the research is conducted, along with citations and a summary of the steps taken. This level of clarity ensures users can understand not only the final answer but also the underlying methodology that led to it.

The time it takes to return a response can vary between 5 to 30 minutes, depending on the complexity of the query. In the future, OpenAI promises that deep research will be able to include embedded images, charts, and other rich media to further enhance its findings.

Capabilities and Limitations

Deep research has been described by OpenAI as capable of performing at the level of a research analyst, capable of processing complex queries and retrieving nuanced insights. For example, a demo query about shifts in the retail industry over the past three years resulted in a detailed, organized response with bullet points and tables outlining the key trends.

However, like any AI, it does come with its limitations. OpenAI warns that the system can "hallucinate" — meaning it may sometimes generate incorrect or fabricated facts. It also may struggle with differentiating between credible sources and rumors, and users should be cautious about the level of certainty in the responses. Despite these challenges, OpenAI’s goal is to continue refining the tool to reduce errors and improve reliability.

Cost and Access

The deep research feature is available to ChatGPT’s Pro users, who pay a $200 monthly fee for access. With this subscription, users are allotted up to 100 queries per month. For those using other subscription tiers like Plus, Team, or Enterprise, access is more limited, but OpenAI plans to roll out higher query limits in the future as the service becomes more cost-effective and computationally efficient.

A New Benchmark for Accuracy

OpenAI’s deep research model has already demonstrated a significant leap in accuracy, surpassing previous benchmarks. In tests conducted using an AI assessment called Humanity’s Last Exam, which includes expert-level questions, the deep research model scored an impressive 26.6% accuracy rate when using browsing and Python tools. This is notably higher than GPT-4’s performance at 3.3%, and even outpaces its next closest model, o3-mini, which scored 13%.

The Future of AI Agents

OpenAI’s deep research feature is part of a broader trend in the AI industry toward creating more autonomous and capable tools. This launch comes closely after the introduction of Operator, a tool designed to complete tasks through a web browser, showcasing the potential for AI to serve as an all-in-one assistant for complex, knowledge-based tasks. Although Google’s Project Mariner—a similar research prototype—is still not available to the public, OpenAI's deep research could be seen as a direct answer to this growing demand for more sophisticated AI agents.

For those willing to invest in the Pro version, deep research is a significant leap forward in AI’s ability to assist with intricate, time-consuming tasks, offering both a glimpse into the future of AI-powered research and a glimpse into the potential of these tools becoming essential for professionals across industries.