An enterprise AI development and operations platform from Microsoft. You pick foundation models from a catalog spanning vendors such as OpenAI, Anthropic, Meta, and xAI, then build AI agents with memory, knowledge, and tool integration through SDKs and APIs for Python, .NET, and other languages. Observability via tracing and evaluation is built in, along with governance features such as permission management and guardrails, and finished agents can be deployed to Microsoft Teams or Microsoft 365 Copilot. It was formerly called Azure AI Foundry and was renamed Microsoft Foundry in 2026, with functionality and existing deployments carried over unchanged.
Key Features
- Multi-vendor model catalog: Choose from OpenAI, Anthropic, Meta, Mistral, DeepSeek, and Microsoft’s own models in a single catalog. The official pricing page describes more than 11,000 models as available, and benchmarking and routing mechanisms help you match a model to each task
- Agent building with Foundry Agent Service: Define agents with memory (retained conversation and context), knowledge, and external tool calls from SDKs and APIs. Hosted agents that run on Foundry’s own infrastructure are supported, as are workflows that coordinate several agents
- External integration via MCP and A2A: Connect to external tools and data sources through the Model Context Protocol (MCP), and expose or call Agent-to-Agent (A2A) endpoints so agents can talk to each other directly. Ready-made tools such as web search and managed MCP servers via connectors are also provided
- Observability and evaluation: Collect execution traces to follow agent behavior, with export to external systems through OpenTelemetry. Traces can be converted into evaluation datasets so agent quality can be measured continuously
- Knowledge grounding with Foundry IQ: A retrieval layer (agentic RAG) built on Azure AI Search lets you ground answers in internal documents
- Governance and enterprise integration: A control plane handles permission management and content safety. Agents you build can be deployed to Microsoft Teams or Microsoft 365 Copilot and placed directly into existing work environments
Pricing
| Category | Price | Details |
|---|---|---|
| Foundry platform | Free | No additional charge for using the portal, browsing the model catalog, or creating agents |
| Foundry Models | Usage-based | Billed on token consumption and inference; unit rates differ by model |
| Foundry Agent Service | Usage-based | No extra charge for creating or running agents; you pay for the model tokens and tool usage consumed |
| Foundry Tools / Foundry IQ | Usage-based | Tool APIs such as OCR, translation, and speech, and Azure AI Search-based retrieval, are billed at their own standard rates |
| Agent Commit Units (ACU) pre-purchase | Contact sales | A one-year commitment plan with tiered discounts: 5% at 20,000 ACUs, 10% at 100,000 ACUs, and 15% at 500,000 ACUs |
Foundry follows a “free platform, pay for what you use” structure rather than a fixed monthly subscription. Per-unit ACU pricing is not published, so a real estimate requires the pricing calculator or a conversation with sales.
Pricing information is current as of August 2026. Check the official pricing page for the latest details.
Pros & Cons
✅ Pros
- Compare and switch between models from OpenAI, Anthropic, Meta, and others in one environment, making single-vendor lock-in easier to avoid
- Model selection, agent building, evaluation, monitoring, and governance live on one platform instead of a patchwork of tools
- Built-in deployment paths to Microsoft Teams and Microsoft 365 Copilot make internal rollout practical
- Support for newer integration standards such as MCP and A2A makes it easy to connect external tools and third-party agents
- The platform itself is free to use, so you can start small and expand
⚠️ Cons
- An Azure account and subscription are prerequisites, which adds setup effort for casual individual use
- Usage-based billing makes costs hard to forecast depending on the mix of models, tools, and retrieval
- The feature surface is broad, and understanding the pieces ─ Models, Agent Service, IQ, Tools, and the control plane ─ carries a learning cost
- Many capabilities are in preview, so operations must keep up with specification changes
- Some pricing (such as ACU unit rates) is unpublished, which makes upfront comparison harder
Comparison with Similar Services
| Criteria | Microsoft Foundry | Amazon Bedrock | Google Vertex AI | OpenAI Platform |
|---|---|---|---|---|
| Provider | Microsoft | Amazon (AWS) | OpenAI | |
| Model choice | Across multiple vendors | Across multiple vendors | Gemini-centric plus external models | Mainly first-party models |
| Agent capabilities | Foundry Agent Service (hosted, multi-agent) | Bedrock AgentCore and similar | Agent Builder and similar | Assistants / Agents API |
| Business tool integration | Deploy to Teams and Microsoft 365 Copilot | Integrated with AWS services | Integrated with Google Workspace and GCP | API-centric (you build the integration) |
| Pricing model | Free platform plus usage-based billing | Usage-based | Usage-based | Usage-based |
All of these are enterprise AI foundations from cloud providers, and their directions are similar. The deciding factor is usually which cloud and which work tools you already use ─ for organizations standardized on Microsoft 365 and Teams, Foundry offers the shortest path.
Who Is It For
- Enterprise IT teams standardized on Microsoft 365 and Teams that want to embed AI agents into business apps
- Developers who want to design systems around comparing models from multiple vendors and using each where it fits
- Teams that want “AI you can actually operate” ─ evaluation, tracing, and permission management included ─ rather than a prototype that stops there
- Practitioners who need answers grounded in internal documents (RAG) and want the retrieval foundation set up as well
- Developers considering integrations with external tools or other agents through MCP and A2A
Summary
Microsoft Foundry brings model selection, agent building, evaluation, monitoring, and governance together into a single enterprise AI platform. Since the rename from Azure AI Foundry, capabilities that support real operations have accumulated: hosted agents, MCP and A2A integration, and turning traces into evaluation datasets. Because the platform itself is free to try, a good starting point is building one small agent and getting a feel for evaluation and tracing. The flip side is that usage-based cost estimation and the sheer number of moving parts are worth understanding up front.