Microsoft Foundry is the control centre for organisations building production AI on Azure. It brings model selection, agent development, grounding, evaluation, deployment and governance into one platform. The value is breadth: teams can move from a prototype to a monitored enterprise service without stitching together a separate model catalogue, safety layer and operations stack.

What stands out

The model catalogue spans more than 11,000 models and includes major commercial and open-source families. Developers can compare models, fine-tune or distil them and use model routing to balance quality and cost. Foundry Agent Service and the Microsoft Agent Framework add managed identity, tools, memory and orchestration for multi-agent systems.

  • A large catalogue covering Microsoft, OpenAI, Anthropic, Meta, Google, xAI and open models
  • Prompt, code and visual development paths for AI apps and agents
  • Grounding through enterprise data, Foundry IQ, connectors, MCP servers and custom APIs
  • Built-in evaluation, tracing, monitoring and content-safety controls
  • Identity and governance integrations with Microsoft Entra, Purview and Defender
  • Deployment across Azure infrastructure with enterprise networking and access controls

Pricing

Foundry does not have one flat subscription price. Billing is consumption based and depends on the models, agent capabilities, storage, search, compute and other Azure services used by a project. Exploring the catalogue is free, but building and deploying requires an Azure subscription. Microsoft also offers an Azure free account for eligible new customers; production buyers should model token usage and supporting services together because the model endpoint is only part of the total cost.

Our take

Foundry is strongest for organisations already invested in Azure or those that need procurement, identity, private networking and governance around several AI models. It can be excessive for a small team that only needs a hosted API. For enterprise programmes, though, the common evaluation and operations layer is a real advantage, especially when teams want to avoid committing every application to one model provider.