Source: Introduction to the Deep Research API (OpenAI Cookbook)

OpenAI cookbook recipe for the o3/o4-mini deep research models: background mode, tools, output items and the clarify-then-rewrite prompt pattern.

Contents

Summary

Shows how to call OpenAI’s deep research models through the Responses API and how to structure the surrounding pipeline.

Key claims

  • Models: o3-deep-research-2025-06-26 for depth and quality, o4-mini-deep-research-2025-06-26 for lower latency.
  • Requires web_search_preview as a data source; code_interpreter is optional for analysis and charts; an MCP tool can expose private documents.
  • Run with background: true because research takes minutes.
  • Output items: reasoning summaries, web_search_call, code_interpreter_call, mcp_call, and a final message whose annotations give title, URL and character span.
  • The API does not ask clarifying questions, so the recipe puts a light model first: one step asks the user a few scoping questions, another rewrites the answers into detailed research instructions.
  • The MCP example sets approvals to never and tells the model not to re-fetch documents it already has.

Feeds

Deep research agents

Suggest changes →