OpenAI Deep Research

Overview

OpenAI has introduced deep research, a new agentic capability within ChatGPT designed to automate complex, multi-step internet research. By synthesizing hundreds of online sources, the tool can complete research tasks in tens of minutes that would typically take a human several hours, producing comprehensive reports with full documentation and citations.

Technical Foundation and Methodology

Deep research is powered by a version of the upcoming OpenAI o3 model optimized specifically for web browsing and data analysis.

Training and Reasoning

Deep research was developed using end-to-end reinforcement learning on challenging browsing and reasoning tasks. It utilizes the same reinforcement learning methods used for OpenAI o1, enabling the model to:

  • Plan and execute multi-step trajectories to locate data.
  • Backtrack and pivot its search strategy in real-time based on encountered information.
  • Browse user-uploaded files and utilize Python tools to plot and iterate on graphs.
  • Cite specific sentences or passages from sources to ensure verifiability.

Performance Benchmarks

The model powering deep research has achieved new state-of-the-art (SOTA) results on several public evaluations:

  • Humanity’s Last Exam: The model achieved 26.6% accuracy on this expert-level test across 100+ subjects, significantly outperforming GPT-4o (3.3%) and OpenAI o1 (9.1%).
  • GAIA: Deep research reached a new SOTA on the GAIA benchmark, which tests reasoning, multi-modal fluency, and tool use. It achieved a 67.36% average score (pass@1) and 72.57% (cons@64).

Key Capabilities and Use Cases

Deep research is designed for intensive knowledge work in fields such as finance, science, policy, and engineering, as well as for high-stakes consumer research (e.g., purchasing cars or appliances).

Functional Differences from GPT-4o

While GPT-4o is optimized for real-time, multimodal conversations, deep research is built for multi-faceted, domain-specific inquiries where depth, detail, and verification are critical. It transforms a quick summary into a documented work product.

Workflow and User Experience

Users select "deep research" in the ChatGPT message composer and provide a query. The process typically takes between 5 and 30 minutes. During this time, a sidebar provides a real-time summary of steps taken and sources used. The final output is a detailed report within the chat.

Access and Iterative Updates

OpenAI is deploying deep research iteratively to manage high compute intensity.

Availability and Limits

As of April 24, 2025, usage limits were updated to include a lightweight version powered by o4-mini to increase efficiency:

  • Pro users: 250 queries per month.
  • Plus, Team, Enterprise, and Edu users: 25 queries per month.
  • Free users: 5 queries per month.

Feature Evolution

  • February 10, 2026: Added connectivity to MCP or apps, the ability to restrict searches to trusted sites, real-time progress tracking, and the ability to interrupt and refine research via follow-up prompts.
  • July 17, 2025: Introduced "agent mode" via the ChatGPT agent, providing access to a visual browser for broader and deeper research.

Limitations and Safety

Despite its capabilities, deep research has known limitations:

  • Accuracy: It can still hallucinate facts or make incorrect inferences, though at a lower rate than previous ChatGPT models.
  • Calibration: The model occasionally struggles to distinguish rumors from authoritative information and may fail to convey uncertainty accurately.
  • Formatting: Minor formatting errors in reports and citations may occur.

Regarding safety, OpenAI's preparedness framework identified the early version of o3 powering deep research as Medium risk. The lab has implemented mitigations specifically for web-browsing risks and continues to monitor the limited release.

Future Roadmap

OpenAI plans to expand deep research's access to specialized data sources, including subscription-based or internal resources. The long-term vision is the integration of deep research (asynchronous investigation) with Operator (real-world action) to enable ChatGPT to execute increasingly sophisticated, end-to-end tasks.

Sources