How OpenAI Built its Groundbreaking Deep Research Product ft. Isa Fulford

Product Overview

• Deep Research is an agentic capability within ChatGPT designed to conduct multi-step online research to solve complex tasks.
• The system operates over a duration of 5 to 30 minutes, synthesizing data into a fully cited, comprehensive report at the level of a research analyst.
• It powers its capabilities with a specialized version of the o3 model, fine-tuned specifically for web browsing and data analysis.

Development & Training Strategy

• Development began by identifying a gap between general progress in math/science/coding models and the need for models trained directly on daily user tasks like online browsing.
• Researchers selected read-only browsing as an initial sandbox to balance utility with constrained safety considerations.
• The team initially built a non-trained demo by directly prompting existing models to validate user interest and visualize product potential.
• Subsequent training involved creating specific reinforcement learning tasks to teach browsing, data analysis, and tool usage.
• Models were granted access to a browser environment (for searching, clicking, scrolling) and code execution tools (for data analysis and graphing).

Use Cases & Demonstrated Capabilities

• Professional Applications: Primary user base includes academics, venture capitalists, and consultants; use cases include analyzing venture capital trends and generating visual data graphs.
• Personal Applications: High utility found in complex shopping and travel recommendations requiring multi-source synthesis.
• Interactive Clarification: The system initiates a "research class" by asking clarifying questions to ensure user specificity before executing the full research task.
• Transparency: Users can view an interleaved "chain of thought" showing the model's planning, reasoning, tool calls (searches vs. Python analysis), and findings in real-time.
• Cross-lingual & Multi-Source Synthesis: Successfully synthesized data from English platforms (e.g., Reddit) and non-English sources (e.g., Korean market reports) to recommend night markets based on proximity and ratings.
• Citation Granularity: Reports include citations that reference specific lines within source documents rather than just the page level.

Technical Insights & Limitations

• The o3 model's strong search performance is attributed to it being trained on the same browsing datasets and tools developed for Deep Research.
• Deep Research represents the "far end" of search duration compared to standard o3 queries, optimized for deep investigation rather than quick answers.
• The system acknowledges current limitations, noting that it can hallucinate, and is actively working to improve reliability.
• One demonstrated graph contained an anomaly (skewed investment data) attributed to the model's inclusion of OpenAI's own investment figures in its analysis.

Future Roadmap

• Integration: Plans to upstream Deep Research tools and capabilities into the main family of reasoning models.
• Private Context: Aiming to integrate private company knowledge and internal sources (like "pay world sources") to enable context-aware research.
• Action-Oriented Agents: The next strategic phase involves moving beyond information synthesis to enabling the model to take autonomous actions based on its findings.