Azure Foundry as a Claude Deployment Target

For organisations with a deep Azure commitment, Anthropic's Claude models in Microsoft Foundry might feel like the obvious choice. Azure is where your infrastructure lives. Identity is already anchored to Microsoft Entra ID. The integration story appears straightforward. But before you proceed, you need to understand a material constraint that separates Foundry from the AWS Bedrock and Google Vertex AI routes, and that constraint is not cosmetic. Anthropic's guarantee that conversation data does not reach Anthropic's systems applies only to Bedrock and Vertex. This guarantee does not yet apply to Azure Foundry. For enterprises with strict "no data to Anthropic" compliance requirements, this is a disqualifier until the guarantee changes. For organisations whose compliance questions live elsewhere, Foundry can still be viable. The route carries genuine advantages: deep Azure integration, Microsoft Entra ID authentication, regional deployment in East US 2 and Sweden Central, and a path to equivalent guarantees announced as coming. This chapter explains the current state, the gaps in available documentation, and when Foundry is the right choice despite the guarantee gap.

The data guarantee gap and why it matters

The boundary between routes is worth understanding in full before any procurement decision. Anthropic's Cowork 3P documentation states the limit precisely. When the inference provider is Vertex or Bedrock, Anthropic guarantees that prompts, responses, and files do not leave Anthropic's systems. This is the "no conversation data sent to Anthropic" statement that appears throughout the Cowork 3P reference. When the inference provider is Azure Foundry or a custom gateway, Anthropic's documentation explicitly states that these guarantees do not apply, and that equivalent guarantees for Foundry are coming in the future.

Read plainly, this means two things. On Bedrock and Vertex, as of 2026-04-22, Anthropic's position is that conversation data does not reach Anthropic. On Azure Foundry and on any gateway route, the same guarantee does not currently exist, and Anthropic's wording is that it is working on parity for Foundry rather than that parity exists now.

The operational consequence is straightforward. If your compliance posture depends on the assertion that no conversation data reaches Anthropic, Bedrock and Vertex are the only two routes that currently support it. Foundry can be a reasonable choice for an organisation whose compliance questions live elsewhere, but it does not carry Anthropic's guarantee in this area today. For a CTO mapping routes to requirements, this boundary is the first filter to apply.

Claude models available in Foundry

Microsoft lists four Claude model families available in Azure AI Foundry. Claude Mythos is a gated research preview intended for evaluation and testing. Claude Opus represents the highest capability tier and supports extended thinking, citations, and adaptive effort parameters that let you shift the quality-to-cost tradeoff. Claude Sonnet offers strong performance on reasoning and coding tasks. Claude Haiku is optimised for speed and operates at the lowest cost. Across all models, Foundry offers feature parity with Bedrock and Vertex, including large context windows, agentic skills, citations, and prompt caching.

One critical detail applies to all Claude deployments on Foundry. Claude uses Global Standard deployment, which means prompts and outputs may be processed outside the deployment region for operational purposes. This is not a regional guarantee. Data at rest remains in the selected Azure region, but inference processing does not. The distinction matters for organisations with strict regional data-processing requirements.

Regional availability and subscription requirements

Claude models on Azure AI Foundry are available for deployment in exactly two regions. East US 2 is the primary region for deployment. Sweden Central is the second option, which provides an alternative for European organisations with specific data residency needs. Both regions must be where your Foundry project or hub is located. You cannot deploy Claude models to other Azure regions even if your organisation uses them for other workloads.

A subscription barrier applies that is worth understanding for procurement. Only Enterprise and MCA-E subscriptions are eligible for Claude models in Foundry. Cloud Solution Provider subscriptions are not supported. Sponsored subscriptions using Azure credits only are not supported. Enterprise Accounts physically located in South Korea are not supported. These restrictions narrow the procurement channels and may require your organisation to adjust its Azure purchasing structure.

Authentication: Entra ID and the enterprise model

Microsoft Entra ID is the recommended authentication mechanism for enterprise deployments of Claude on Foundry. Entra ID supports keyless authentication via DefaultAzureCredential, which means your users can authenticate through their existing Microsoft identity without storing long-lived API keys. This integration is straightforward if your device management platform already delivers Entra ID configurations to your managed machines.

API keys remain as a fallback option for scenarios where Entra ID is not available or for single-machine evaluation. However, a notable restriction applies to the Mythos preview. Mythos requires Microsoft Entra ID authentication and does not support API key authentication. If you intend to trial Mythos, your evaluation environment must support Entra ID integration.

Data residency and the Global Standard caveat

Claude on Foundry uses the Global Standard deployment model. This designation has a specific meaning that deserves clarity. Microsoft's documentation states explicitly that "prompts and outputs may be processed outside of your region, for operational purposes. Operational purposes include performance and capacity management." Data at rest remains in the designated Azure geography. Data in flight during inference does not.

This distinction is material. An organisation with genuine regional lock requirements, where prompts and outputs must be processed within a specific Azure region without exception, will not find that guarantee in the Global Standard deployment. The tradeoff is that Global Standard provides better performance and capacity utilisation across Anthropic's infrastructure. If your requirements demand absolute regional processing guarantees, you would need to look at other Azure options or a different vendor altogether.

Pricing and feature parity

Claude pricing in Azure AI Foundry is not published on Microsoft's public pricing pages. Direct quotes from your Microsoft account manager are required to determine the cost-per-token for each model. Expect conversations around per-million input and output token costs, with volume discounts negotiable. Feature parity with Bedrock and Vertex means you have access to all Claude capabilities, including the effort parameter, which lets you adjust inference quality and cost on a per-request basis. Fine-tuning is not available for Claude models on Foundry.

The lack of public pricing is typical of premium enterprise offerings but adds complexity to procurement. Budget estimates require direct vendor negotiation rather than forward planning from published rates. Build in procurement timeline accordingly.

Documentation gaps and what to ask your account manager

Three material gaps exist in Microsoft's published documentation for Claude on Foundry. These gaps matter for procurement and should be clarified before signature.

First, pricing transparency. Token costs and volume discounts are not publicly listed. Your account manager can provide the information, but the absence of public pricing makes it difficult to budget and compare Foundry against Bedrock or Vertex without direct conversation. Request explicit per-token costs for each model, volume tiers, and any premium charges for advanced features.

Second, customer-managed keys support. CMK is essential for regulated enterprises that must control encryption key lifecycle. Microsoft's documentation does not explicitly state whether CMK is supported for Claude deployments. This is a critical gap that must be closed before procurement. Ask your account manager whether CMK is available and, if so, what the implementation looks like.

Third, EU Data Boundary specifics. Organisations operating under EU Data Boundary requirements need clarity on whether Sweden Central deployments qualify. The Sweden Central region is physically located in Sweden, but the Global Standard deployment model allows processing outside the region "for operational purposes". This creates ambiguity about whether EU Data Boundary requirements are met. Request explicit written confirmation from your account manager about EU Data Boundary coverage, or escalate to Anthropic for clarity.

When Foundry is the right choice

Foundry is the right choice when your organisation is Azure-first or Azure-committed and when either your compliance constraints allow the current guarantee gap or your organisation is willing to wait for equivalent guarantees. Foundry fits if your DevOps teams are already fluent in Azure governance, if Entra ID is your identity standard, and if regional processing in East US 2 or Sweden Central aligns with your infrastructure footprint.

Foundry does not fit when your organisation has a strict "no data to Anthropic" requirement that cannot be waived, when your infrastructure is locked to a different cloud provider, or when regional processing guarantees are non-negotiable. In those cases, Bedrock on AWS or Vertex AI on Google Cloud are the routes that carry the guarantees you need.

What this chapter does not cover

This chapter focuses specifically on Foundry. For broader context on the architecture of Cowork 3P, deployment mechanisms, and security controls that apply across all routes, see Chapter 1. For the consolidated controls matrix comparing Foundry against Bedrock and Vertex on telemetry, egress, sandbox isolation, and identity, see Chapter 8. For governance implications of Azure Policy integration and MDM configuration specifics, see Chapter 9.

Primary sources

• Microsoft. Claude in Azure AI Foundry setup and deployment guide. Retrieved 22 April 2026.
• Microsoft. Azure AI Foundry model regional availability reference. Retrieved 22 April 2026.
• Microsoft. Data privacy and residency for Claude models in Azure AI Foundry. Retrieved 22 April 2026.
• Anthropic. Cowork on 3P — Overview. Retrieved 22 April 2026.

Nothing in this article is legal advice. It names regulatory frameworks and describes how each deployment route affects compliance posture. Compliance interpretation for your specific regulatory context, jurisdiction, and client contracts must be reviewed with qualified legal counsel. Verify current Anthropic documentation before making a procurement decision.