OpenAI has launched a groundbreaking feature in ChatGPT called Deep Research, poised to transform how we tackle complex research tasks. This powerful new capability empowers ChatGPT to conduct in-depth, multi-step internet investigations, generating comprehensive, analyst-level reports in a fraction of the time it would take a human expert.
Deep Research: An Autonomous Research Agent
Deep Research functions as an autonomous agent, independently searching, analyzing, and synthesizing information from hundreds of online sources. Powered by a version of OpenAI’s upcoming o3 model, optimized for web browsing and data analysis, it can process vast amounts of text, images, and PDFs, adapting its approach as it encounters new information. OpenAI views this knowledge synthesis capability as a crucial step towards Artificial General Intelligence (AGI).
Addressing the Need for Efficient, Reliable Research
OpenAI developed Deep Research to serve professionals in fields like finance, science, policy, and engineering who require thorough and accurate research. It also caters to consumers seeking highly personalized recommendations for significant purchases. Critically, every Deep Research output is fully documented with citations and a summary of its reasoning, ensuring transparency and verifiability. The tool excels at uncovering niche information, saving users countless hours of manual research.
How Deep Research Works its Magic
Trained using end-to-end reinforcement learning on challenging browsing and reasoning tasks, Deep Research plans and executes multi-step research trajectories, dynamically adjusting its approach based on incoming data. It can browse user-uploaded files, generate graphs using Python, and embed images and graphs from websites into its responses. Transparency is paramount, with the model citing specific sentences or passages from its sources.
Impressively, on “Humanity’s Last Exam,” a rigorous AI evaluation, Deep Research achieved a 26.6% accuracy, significantly outperforming other models like GPT-4o and o3-mini. It also set a new benchmark on the GAIA public leaderboard for real-world problem-solving.
Using Deep Research in ChatGPT
Accessing Deep Research is simple: select the “Deep Research” option in ChatGPT’s message composer and enter your query. Users can request anything from competitive analyses to personalized product recommendations. File and spreadsheet attachments provide added context. A sidebar tracks the research progress and sources used. Research typically takes 5-30 minutes, allowing users to multitask. The final report is delivered within the chat, with plans to add embedded images and data visualizations soon.
Deep Research vs. GPT-4: A Key Distinction
While GPT-4 excels in real-time conversations, Deep Research is designed for in-depth, domain-specific inquiries requiring extensive research. For example, instead of simply stating a number, Deep Research provides a comprehensive analysis of the average retirement age for NFL kickers, including statistical context, examples, and influencing factors.
Performance and Limitations
Deep Research has achieved state-of-the-art performance on the GAIA benchmark. However, it’s not without limitations. It may occasionally hallucinate facts or make incorrect inferences, although its error rate is lower than previous ChatGPT models. It can also struggle with source credibility and may not always convey uncertainty effectively. Minor formatting issues may also occur. OpenAI expects these issues to improve over time.
Availability and Future Directions
Deep Research is currently available on ChatGPT’s web version for Pro users, with rollout to mobile and desktop apps, Plus, Team, and Enterprise users, and expansion to other regions planned. Due to its resource intensity, query limits are in place, but OpenAI is developing a faster, more cost-effective version with higher limits.
Looking ahead, OpenAI envisions integrating Deep Research with Operator, enabling ChatGPT to perform even more sophisticated tasks by combining online research with real-world actions. This development promises to further revolutionize how we interact with and leverage AI for complex problem-solving.