Microsoft Foundry in Practice

Deploying Claude Models in Foundry: Versions, Aliases, and Updates

On Microsoft Foundry you don't call a model directly — you call a deployment you created and named. Getting the deployment step right saves confusion for every team that builds on it later.

Claude 3P 101 · Updated July 2026 · Unofficial guide

Microsoft Foundry organizes Claude access in a two-level hierarchy. A Foundry resource holds the security and billing configuration — keys, networking, the Azure Marketplace subscription. Inside the resource you create one or more deployments, and each deployment is a named instance of a specific Claude model. Your applications call the deployment, not the model catalog. This article walks through creating one, choosing its "model version" (which on Foundry means something unusual), and handling updates over time.

Before you deploy: prerequisites

Claude models are third-party Azure Marketplace offerings from Anthropic, so a few conditions apply. You need a paid Azure subscription with a pay-as-you-go billing method — free trials, student subscriptions, and credit-only subscriptions are not supported for Claude. Unsupported types also include Enterprise Accounts in South Korea, Cloud Solution Provider subscriptions, and sponsored subscriptions using only Azure credits. On the access side you need the Foundry portal, Azure Marketplace access, and the Contributor or Owner role on the resource group. If you haven't created the resource yet, start with setting up a Foundry resource.

The deploy flow in the portal

In the new Foundry portal the path is: Discover > Models, select a Claude model, then Deploy > Custom settings. You'll accept the Azure Marketplace terms, then configure three things:

1. Deployment name. It defaults to the model ID (Foundry uses bare first-party IDs like claude-opus-4-8 or claude-sonnet-5, with no anthropic. prefix). You can customize it — but it cannot be changed after creation, and it is exactly the value your code passes in the model parameter of API requests. Many teams keep the default so the name tells you what's behind it; if you use a role-based alias like chat-primary, document the mapping.

2. Region scope. Global or Data Zone — see choosing an Azure region for Foundry for the trade-offs.

3. Model version — really a hosting choice. On Foundry, "version" does not mean a model snapshot. Claude models come in two hosted flavors: version 1 = Hosted on Anthropic infrastructure (runs on Anthropic's infrastructure, outside Azure) and version 2 = Hosted on Azure (runs on Azure end-to-end, Generally Available). Choosing Deploy > Default settings auto-selects Hosted on Azure. Both options are Anthropic-operated services; Anthropic recommends Hosted on Azure for most workloads, and Hosted on Anthropic when you need features or models not yet hosted on Azure.

Rule of thumb: not every model exists in both flavors. Hosted on Azure (GA) offers claude-opus-4-8, claude-sonnet-5, and claude-haiku-4-5. Hosted on Anthropic additionally offers claude-fable-5 (Preview) and older generations such as claude-opus-4-7, claude-opus-4-6, claude-sonnet-4-6, and claude-sonnet-4-5.

After deployment: endpoint and key

Open Build > Models, select your deployment, and the Details tab shows the Target URI (your endpoint, shaped like https://<resource-name>.services.ai.azure.com/anthropic) and the Key. From there, your first API call is a normal Claude Messages request with the deployment name as the model. If you'd rather script the whole thing, Microsoft's "Claude on Foundry starter kit" (github.com/Azure-Samples/claude) provisions the account, project, and deployments with a single azd up using Bicep or Terraform, wired for Entra ID with no API keys.

Updates, deprecations, and staying current

Because deployment names are immutable and pinned to a model, "updating" means creating a new deployment for the newer model and cutting traffic over — a useful forcing function for testing before switching. Foundry follows the Claude API model lifecycle schedule: for example, Claude Opus 4.1 on Foundry is deprecated and retires August 5, 2026. Watch the lifecycle notes so a retirement doesn't surprise a production deployment. Also note that feature support differs by hosting flavor: structured outputs, server-side tools, the MCP connector, Agent Skills, and the Files API are only available on Hosted on Anthropic deployments — sending such requests to an Azure-hosted deployment returns 400 Bad Request by design. And some of Foundry's Claude features are still in beta; see the beta caveats article before promising features to stakeholders.

Where to go next

Next steps: grab credentials in managing API keys in Foundry or skip keys entirely with Entra ID authentication, then wire up the Python SDK. The platform overview puts Foundry in context.

Sources