How to Use Claude AI on Azure | Foundry, Pricing, and Differences

Three Paths for Using Claude AI on Azure

There are currently three official paths for calling Claude from an Azure environment. Based on official announcements from Microsoft and Anthropic (source), these are: a serverless API via Microsoft Foundry, Azure Databricks' Mosaic AI Model Serving, and a path where end users interact directly with Claude embedded in Microsoft 365 Copilot.

The first path, via Foundry, is for developers — it calls the Anthropic SDK's Messages API through Azure's authentication, billing, and monitoring layer. The second, via Databricks, is designed for combining with a data platform, allowing you to pass data from Spark or Unity Catalog directly to Claude. The third, via Copilot, is for business users — Excel, Word, and the Researcher agent call Claude internally, so users can work with it without ever touching an API or key.

All three paths operate entirely within your Azure subscription's billing scope, eliminating the need for a separate contract with Anthropic. Because they qualify for MACC (Microsoft Azure Consumption Commitment), companies that already have an Azure commitment can try Claude with zero new procurement — which is the biggest advantage (source).

Steps to Call Claude via Microsoft Foundry

Microsoft Foundry is Azure's model catalog, and Claude was registered as an official Anthropic partner on November 18, 2025. The workflow is five steps, and following the official Microsoft Learn documentation (source) will get your first request through within 30 minutes.

1. Create a Microsoft Foundry project in East US2 or Sweden Central from the Azure portal.
2. Accept the Anthropic Claude offer on Azure Marketplace and assign Contributor or Owner role to the target resource group.
3. Select a model such as claude-sonnet-4-6 from the Foundry portal's model catalog and deploy it via Global Standard deployment.
4. Note the base URL (https://<resource-name>.services.ai.azure.com/anthropic) and deployment name shown on the deployment details page.
5. Pass the base URL and credentials to the Anthropic SDK's AnthropicFoundry client and call the Messages API.

The SDK officially supports Python, TypeScript, and C#, and you can also call the same endpoint via REST (cURL). Sample code for each language is available on Microsoft Learn, so existing Anthropic SDK users can migrate to the Azure path with just a few lines of changes.

Supported Models and Regions

Foundry's Claude catalog supports 9 models as of May 2026. These include five generations of Opus (4.8 / 4.7 / 4.6 / 4.5 / 4.1), Sonnet 4.6 / 4.5, Haiku 4.5, and one research preview — Mythos Preview, available by invitation for defensive cybersecurity use (source).

Here is a summary of the primary models, their use cases, and limits:

Claude Opus 4.8 (Foundry preview)

The latest flagship. With a 1M-token context window and 128K output limit, it is ideal for enterprise workflows, autonomous coding, financial analysis, and cybersecurity. Adaptive thinking mode automatically adjusts reasoning time based on task difficulty.

Claude Sonnet 4.6 (Foundry preview)

The standard model. It shares the same 1M / 128K window as Opus while offering a balance of speed and cost, making it the top candidate for embedding in applications. Both adaptive thinking and max effort modes are available.

Claude Haiku 4.5 (Foundry preview)

The lightweight model. It delivers "near-frontier" performance for coding and agent use cases while operating at the speed and per-token cost needed to handle free features or large numbers of sub-agents.

Available regions are limited to two locations: East US2 and Sweden Central, and Japanese regions are not supported as of May 2026. For EU customers, there are ongoing discussions on Microsoft Q&A about "timelines for migrating to Azure EU infrastructure," and companies with data sovereignty requirements should plan to operate in Sweden Central for the time being.

Authentication | Entra ID vs. API Key

Claude on Foundry authenticates via either Microsoft Entra ID (formerly Azure AD) or an API key. Entra ID is recommended for enterprise use — by using the DefaultAzureCredential class, you can share the same code across development, staging, and production environments while automatically retrieving credentials from each environment's environment variables or Managed Identity.

API keys are intended for validation and proof-of-concept use; simply pass the apiKey when creating the client and it will work immediately. However, Mythos Preview does not support API keys and requires Entra ID, so if you have research preview use cases, it is safer to set up Entra ID integration from the start.

Key settings to watch for during authentication:

1. The scope for Entra ID authentication is fixed at https://ai.azure.com/.default
2. Assign Contributor or Owner to the resource group and Cognitive Services User to the resource
3. The base URL format is https://<resource-name>.services.ai.azure.com/anthropic; append /v1/messages for the Messages target URI
4. Specify the API version header as anthropic-version: 2023-06-01

A 403 error almost always indicates a missing RBAC setting; a 404 is usually a typo in the endpoint URL or deployment name. Microsoft Learn's troubleshooting table summarizes the resolution steps for common errors.

How to Use Claude via Azure Databricks

When combining with a data platform, you can call Claude through Azure Databricks' Mosaic AI Model Serving. Using the external models feature lets you manage the Anthropic official API through Databricks' Unity Catalog, enabling direct calls to Claude from SQL queries or Spark jobs (source).

As of May 2026, Anthropic Foundation Model APIs (pay-per-token mode) have not yet reached full tenant availability and are limited to specific regions. Before adopting in production, it is safest to confirm with Microsoft support whether your Databricks workspace is in a supported region.

Using Claude in a RAG configuration sourced from internal documents on a data lakehouse, or generating summaries with Claude at the final stage of a bronze/silver/gold pipeline — Databricks is the shortest path for these configurations. A simple decision rule: choose Foundry or Databricks based on where your data lives.

Claude Integrated into Microsoft 365 Copilot

The third path for business users is Claude's integration into Microsoft 365 Copilot. According to Anthropic's official announcement, Claude is embedded as an option in the Copilot Researcher agent and Excel's Agent Mode, and users can switch between Sonnet and Opus simply by selecting "Claude" from the model selection menu (source).

In the Researcher agent, Claude handles web search, internal data search, and PDF analysis end-to-end. In Excel Agent Mode, formula generation, data formatting, and analysis prompt execution can be offloaded to Claude. The key advantage over other paths is that users can access Claude's processing power within their Microsoft 365 license without ever dealing with API keys or project IDs.

However, since the Copilot path means model versions are managed by Microsoft, it is not suited for use cases where you want to try the latest Opus 4.8 the day it releases. If you need fine-grained prompt control, thinking parameter tuning, or long-context utilization, choose the Foundry path.

Differences from Azure OpenAI Service

Azure OpenAI Service and Foundry's Claude are sibling services that can both be called from the same Azure billing scope. The difference is "who developed the model" — Azure OpenAI Service provides GPT-5 / GPT-4.1 models, while Foundry's Claude provides Anthropic models.

Here is a summary of the key differences:

If you have Japan data residency requirements and must use the East Japan region, choose Azure OpenAI. If your primary use cases are long-form reasoning or Computer Use agents, choose Foundry's Claude. Running both in parallel and routing by task is also realistic on Azure.

Pricing and Rate Limits | Enterprise Subscription Required

First, let's clarify some terms. MACC (Microsoft Azure Consumption Commitment) refers to the pre-committed quota you secure by committing to Azure usage in advance; Foundry's Claude charges draw from this quota. MCA-E (Microsoft Customer Agreement - Enterprise) is the direct enterprise contract with Microsoft; standard pay-as-you-go subscriptions and individual plans are not eligible. For PoC purposes, use your company's Azure Enterprise tenant.

Foundry's Claude is currently limited to Enterprise subscriptions or MCA-E (Microsoft Customer Agreement - Enterprise) contracts. Individual developers on pay-as-you-go or Visual Studio subscriber benefits cannot use it, so PoC work should be done on a company Azure tenant.

Enterprise rate limits for primary models are as follows:

• Claude Opus 4.8 / 4.7 / 4.6 / 4.5 / 4.1: 2,000 RPM / 2,000,000 TPM
• Claude Sonnet 4.6: 2,000 RPM / 2,000,000 TPM
• Claude Sonnet 4.5: 4,000 RPM / 2,000,000 TPM
• Claude Haiku 4.5: 4,000 RPM / 4,000,000 TPM

If you exceed these limits, submit a quota increase request via Microsoft's form. Token pricing is on par with direct Anthropic contracts, and since charges draw from Azure's MACC commitment, companies already committed to Azure can effectively use Claude at a discount.

Decision Framework for Choosing a Path

Choosing among the three paths is clearest when you evaluate along three axes: engineering setup, where your data lives, and who your users are.

1. An engineering organization that writes its own APIs → Call Microsoft Foundry directly (Anthropic SDK + Entra ID)
2. Already using Databricks as your data platform → Azure Databricks Mosaic AI Model Serving
3. Business users accessing from Excel / Word → Via Microsoft 365 Copilot
4. Japan data residency requirements (domestic storage of personal or confidential data) → Use Azure OpenAI Service's East Japan / West Japan regions; for long-form tasks requiring Claude, either process cross-border via API or confirm with legal whether Sweden Central is acceptable before adopting

When in doubt, start by deploying Sonnet 4.6 to Foundry and making one request with the Python SDK. Once Foundry is working, you can expand step by step within the same Azure tenant — Databricks integration, Copilot integration, parallel use with Azure OpenAI, and multi-LLM routing all consolidate onto the same invoice. Options for combining Azure and Claude continue to expand in 2026, and now is the peak time to incorporate "Claude on Azure" into your internal AI infrastructure roadmap.