Claude in Azure Foundry Goes GA: What the New Hosting Architecture Means for Enterprise Routing

Anthropic and Microsoft shipped a significant infrastructure upgrade today: Claude in Microsoft Foundry is now generally available, and the architecture has changed in ways that matter for teams deciding how to route to Claude in enterprise environments.

What changed

The Foundry Preview that existed before today ran Claude models through Anthropic's own infrastructure, billed through the Azure Marketplace as a passthrough. The GA launch introduces a second, distinct deployment type — Hosted on Azure — where inference actually runs on Anthropic-operated services within Azure, on NVIDIA GB300 Blackwell Ultra hardware connected by InfiniBand networking.

This creates two routing paths for Claude through Foundry:

The default recommendation for most workloads is now Hosted on Azure.

The billing and procurement changes

Claude in Foundry now bills in Claude Consumption Units (CCU) — a single, consolidated line on your Azure bill, with per-model detail available inside Foundry. For organizations with a Microsoft Enterprise Agreement (EA), Claude usage qualifies for Microsoft Azure Consumption Commitment (MACC) drawdown, meaning Claude calls can draw down an existing Azure spend commitment rather than requiring a separate Anthropic contract.

For procurement teams at enterprises that already have large Azure commitments, this matters more than model benchmarks. It collapses two vendor relationships (Azure and Anthropic) into one billing line, using commitments the enterprise has already negotiated.

What this changes for Claude azure foundry enterprise routing policy

Authentication: Claude in Foundry now uses Microsoft Entra ID for authentication. Teams that previously authenticated with Anthropic API keys can now use Azure role-based access controls (RBAC), apply existing identity governance policies, and audit Claude access through the same tooling used for other Azure services. For organizations with strict identity governance, this eliminates a separate credential lifecycle for the Anthropic key.

Data residency: The US Data Zone Standard deployment type keeps inference within US Azure regions and US data processing boundaries. For teams with US data residency requirements — regulated industries, government contractors, financial services — this provides a path to Claude that stays within their existing Azure compliance envelope. The Global Standard type remains available for teams without strict geo requirements.

Zero data retention: Prompts and completions sent through Foundry are not retained by Anthropic after the API call completes, when ZDR is enabled. This is configurable for high-sensitivity workloads.

Network path: Because Claude now runs inside Azure, organizations can use Azure private networking, virtual networks, and existing network security group policies to route Foundry traffic without leaving the Azure perimeter. Versus calling the Anthropic API directly, this can eliminate a class of egress and external-dependency concerns for security-sensitive workloads.

The two-path routing decision

The GA release formalizes a choice that now has operational consequences:

Choose Hosted on Azure when:

• Your organization has MACC commitments you want to draw down
• You need a US data zone or data residency enforcement
• Your identity governance requires Entra ID authentication for all API calls
• Your networking policy restricts calls to Azure-perimeter endpoints
• You are running Opus 4.8 or Haiku 4.5 (the currently supported models)

Choose Hosted on Anthropic (the existing Foundry Preview mode) when:

• You need models not yet hosted on Azure (Fable 5, Mythos 5, Sonnet variants)
• You need API features not yet available on the Azure-hosted path (check the docs for the current matrix)
• You have workloads that were already running in Foundry Preview and need time to validate the new path

Anthropic has stated that feature and model parity between the two modes is a goal over time.

What to check in your current setup

If your team is already using Claude through Foundry (the Preview), your traffic is on the Hosted on Anthropic path. The GA release doesn't migrate existing deployments; you choose the hosting mode when configuring a deployment. If you want to move to the Hosted on Azure path, you need to create a new deployment in Foundry with the Azure hosting option selected.

SDK support: Python, TypeScript, C#, Java, and PHP SDKs support Foundry. Go and Ruby SDKs do not currently support Foundry — teams on those runtimes need to make direct HTTP calls or wait for SDK support.

Pricing: Claude usage in Foundry is now priced in CCU through the Azure Marketplace. Check the Claude pricing page for current CCU rates. The per-model rate structure is the same as Anthropic direct API pricing; the difference is the billing envelope and MACC eligibility.

The broader routing picture

This GA launch changes the decision matrix for enterprise teams evaluating Claude routing options. Before today, the practical options were: Anthropic API directly, Amazon Bedrock, Google Cloud Vertex AI, or the Foundry Preview (Anthropic-hosted). The GA adds a meaningful fourth enterprise path — Anthropic on Azure — that combines frontier model quality with Azure-native procurement, identity, and compliance controls.

For multi-provider routing setups, the Foundry models (Opus 4.8, Haiku 4.5 on Azure-hosted path) are now viable as primary endpoints for enterprise teams with Azure-first infrastructure. Teams that previously routed Claude calls direct to the Anthropic API to avoid a separate cloud contract now have an Azure-native alternative that maps onto existing commitments.

For routing teams managing cost attribution, the CCU billing line provides per-model granularity inside the Azure cost management tooling — which integrates with existing tagging, budgets, and alert policies — rather than requiring a separate Anthropic billing view.

More details on setup, deployment types, and the API endpoint are in the Anthropic Foundry docs.