Deep Research launches in Azure AI

Microsoft has announced the public preview of Deep Research, a new feature within the Azure AI Foundry Agent Service. This offering delivers OpenAI’s advanced agentic research capabilities as an API and SDK solution, integrated directly with Azure’s enterprise-grade agentic platform.

Deep Research Launches in Azure AI Foundry Agent Service: Enterprise-Scale Web Research Automation Now Available in Preview

Through Deep Research, developers can now build intelligent agents that perform in-depth planning, analysis, and synthesis of web-based information. The service is designed to automate complex research tasks, generate transparent and auditable outputs, and integrate multi-step workflows across Azure AI Foundry’s ecosystem of tools and agents.

Extending Generative AI to Enterprise Knowledge Work

Tools like ChatGPT Deep Research and Researcher in Microsoft 365 Copilot have already brought generative AI into individual and team-based productivity workflows. These tools are transforming how millions of users handle research and documentation.

Now, enterprises aiming to embed deep research within business applications, automate multi-step processes, and manage knowledge at scale can leverage Deep Research as a programmable and composable service. Azure AI Foundry provides a foundation for such orchestration, enabling seamless integration with internal data and systems.

Core Capabilities of Deep Research

The Deep Research offering is built for developers ready to move beyond traditional chat-based interfaces. As a composable agent available via API and SDK, it empowers teams to:

• Automate web-scale research grounded in Bing Search, ensuring all insights are traceable and source-supported.
• Build programmable agents that can be invoked by applications, workflows, or other agents—transforming research into a production-ready service.
• Compose sophisticated workflows using Deep Research in conjunction with Azure Logic Apps, Azure Functions, and other Foundry Agent Service connectors for automating tasks such as reporting or notifications.
• Maintain enterprise governance with security, compliance, and observability features that offer full control over how research processes are executed and tracked.
   

Unlike prebuilt chat assistants, Deep Research evolves alongside an enterprise’s needs, offering automation, extensibility, and the potential for deeper integration with proprietary data sources as future support expands.

How Deep Research Works: Under-the-Hood

Deep Research has been engineered for flexibility, transparency, and composability, enabling it to meet the complex demands of modern business research.

At the center of the system is the o3-deep-research model. This orchestrates a multi-step pipeline that integrates grounding with Bing Search and the latest OpenAI models:

• Intent Clarification and Task Scoping: Upon receiving a research query, the agent uses GPT-series models (including GPT-4o and GPT-4.1) to refine the user’s intent and define a precise research task. This step ensures outputs are contextually relevant and useful for the given business scenario.
• Web Grounding with Bing Search: After scoping, the agent initiates a secure search using Bing, collecting a curated, high-quality, and recent dataset to serve as a foundation. This mitigates hallucination and enhances reliability.
• Research Execution: The o3-deep-research model analyzes and synthesizes the discovered data. It processes information step-by-step, adjusts to new insights, and constructs a comprehensive, nuanced output rather than performing mere summarization.
• Auditability and Compliance: Every response includes a structured report containing citations, the reasoning process, and any clarifications asked during the session. This makes the results fully transparent and auditable—an essential feature for industries with regulatory requirements.
• Integration and Workflow Composition: As an API-first solution, Deep Research can be called from virtually anywhere—business apps, workflow tools, portals, or within multi-agent systems. For instance, one agent might conduct a web analysis, another might create a presentation with Azure Functions, and a third could distribute it via email using Azure Logic Apps.
   

This composability allows organizations to transform research from a manual task into a scalable component of digital transformation and enterprise intelligence.

Use Cases and Early Adoption

Early adopter organizations are already exploring how to apply Deep Research agents in high-value scenarios such as market analysis, competitive benchmarking, large-scale analytics, and compliance reporting.

Pricing Model

Deep Research’s o3-deep-research model is priced as follows:

• Input tokens: $10.00 per 1 million tokens
• Cached input tokens: $2.50 per 1 million tokens
• Output tokens: $40.00 per 1 million tokens
   

Search context tokens are billed according to the input token pricing of the model in use. Additional charges apply for the Bing Search grounding service and GPT models used during question clarification.

Availability and Getting Started

Deep Research is currently in limited public preview for Azure AI Foundry Agent Service customers. Interested users can:

• Sign up for early access through the preview program.
• Review official documentation to understand feature implementation.
• Explore learning modules to build their first research agents.
• Begin creating custom research agents directly in Azure AI Foundry.
   

Microsoft has indicated that more updates are on the horizon, including customer success stories, feature enhancements, and broader availability, as the company continues to advance enterprise AI capabilities.

Build your Deep Research agent in Azure AI Foundry