Claude Opus 4.8 and Haiku 4.5 Are Now Generally Available in Microsoft Foundry on Azure

Update, 1 July 2026: Claude in Microsoft Foundry is now GA on Azure with Azure-native identity, billing, and US data zone

The original article covered model availability: Opus 4.8 and Haiku 4.5 going GA in Microsoft Foundry. This update adds the infrastructure layer that now sits underneath those models.

As of today, Claude in Microsoft Foundry is generally available hosted on Azure, running inside a customer’s Azure environment rather than on Anthropic infrastructure. This is a distinct offering from the Anthropic-hosted Foundry Preview that the original article described.

The practical difference matters for enterprise procurement. The Azure-hosted path gives teams Microsoft Entra ID authentication with Azure RBAC, billing through the Azure Marketplace denominated in Claude Consumption Units (eligible for Microsoft Azure Consumption Commitment drawdown), and data processing scoped within Azure. An optional US data zone deployment satisfies data residency requirements for regulated industries. Prompts and completions stay within Azure; only usage metadata and content flagged by safety systems egresses to Anthropic.

Opus 4.8 and Haiku 4.5 are available at launch via the Messages API, with prompt caching and extended thinking supported. Two hosting options now exist: hosted on Azure (described above) and hosted on Anthropic, the prior path, which retains broader feature coverage while Anthropic works toward parity. The full feature delta is documented on the Anthropic platform. Free trial, student, and sponsored Azure subscriptions are not supported.

Claude is now generally available in Microsoft Foundry. Specifically, Claude Opus 4.8 and Claude Haiku 4.5 are live on Azure today, running on NVIDIA GB300 Blackwell Ultra GPUs, with full Azure-native identity, billing, governance, and a US data zone option built in.

This also makes Azure the only cloud platform where you can access both OpenAI and Anthropic frontier models under one roof, on one bill, through one set of controls.

What’s Actually Available

The GA release brings two models into Microsoft Foundry via the Messages API:

• Claude Opus 4.8 — the higher-capability model, suited to complex reasoning, coding, and agentic tasks, with a 200k-token context window
• Claude Haiku 4.5 — the faster, more efficient option for high-volume or latency-sensitive workloads, also at 200k tokens

Both support prompt caching, extended thinking, and tool streaming. The C#, Java, PHP, Python, and TypeScript SDKs all work with Foundry. Go and Ruby do not currently.

For deployments, Foundry uses a two-level structure: you create a resource (which holds your security and billing config), then create model deployments within it. Global Standard deployments are available in East US2 and Sweden Central. Opus 4.8 on Azure also supports Data Zone Standard (US) deployment for teams with data residency requirements.

The Hardware Behind It

Claude in Foundry runs on NVIDIA GB300 NVL72 systems, connected via NVIDIA Quantum-X800 InfiniBand networking. The GB300 NVL72 is a rack-scale, fully liquid-cooled system combining 72 Blackwell Ultra GPUs and 36 Grace CPUs, with 37TB of fast memory, 130TB/s of NVLink bandwidth, and FP4 Tensor Core performance up to 1,440 petaflops with sparsity.

That level of infrastructure matters for inference workloads at scale. Bolt, one of the early customers, put it plainly: “Running Anthropic’s models on Azure has given us the sustained throughput and reliability our enterprise customers expect.” Momentic reports serving millions of tokens per minute through Foundry.

This hardware partnership has significant financial weight behind it. Anthropic has committed to purchasing $30 billion in Azure compute capacity, with additional capacity up to one gigawatt. NVIDIA and Microsoft have each committed investment in Anthropic of up to $10 billion and $5 billion respectively. The GB300 deployment is a direct result of that three-way arrangement, first announced in November 2025.

What This Means for Enterprise Teams

The practical value here is less about the models themselves and more about how they slot into enterprise operations. Most AI projects inside large organisations do not stall because of model quality. They stall because of procurement, access controls, compliance sign-off, and data handling requirements.

Claude in Foundry addresses most of that directly:

Identity and access: Claude authenticates through Microsoft Entra ID, so your existing role-based access controls and governance policies apply without any additional configuration.

Billing: Claude usage is billed in Claude Consumption Units (CCU) and shows up as a consolidated line on your Azure invoice. For customers with a Microsoft Enterprise Agreement, Claude usage draws down your existing Microsoft Azure commitment (MACC). No separate Anthropic billing account, no new procurement relationship to set up.

Compliance: The service inherits Azure’s compliance certifications, including HIPAA, SOC 2, and FedRAMP. Anthropic also brings its own safety layer, and zero data retention is available for high-sensitivity workloads, meaning prompts and completions are not stored after the API call completes.

Data residency: The US data zone option lets teams with specific data residency requirements keep inference processing within US boundaries.

Two Ways to Run Claude in Foundry

There are two hosting paths, and the distinction matters:

Hosted on Azure is the option for teams where running inside their Azure environment is a hard requirement. You get Azure authentication, billing integration, governance controls, and the US data zone. This is the new GA path.

Hosted on Anthropic (previously the Foundry Preview) is still available when you need the full set of API features or a model not yet available on Azure. It runs on Anthropic’s infrastructure but is still accessible through the Foundry interface.

The right choice depends on your compliance requirements and which models or features you need. For most regulated enterprise workloads, Hosted on Azure is the natural default.

The Bigger Picture for Azure

Azure’s position here is notable. AWS hosts Claude on Bedrock, and Google Cloud offers it through Vertex AI, so Claude’s presence across clouds is not new. What is new is Azure being the only platform offering both OpenAI and Anthropic frontier models natively, with the same billing and governance layer applied to both.

For enterprise buyers who have already standardised on Azure, this removes a meaningful friction point. You no longer have to choose between model quality and operational simplicity, or manage a separate external API relationship for Claude alongside your existing Azure footprint.

If you’re building agentic systems, the combination of GB300 throughput and Azure-native controls is particularly relevant. Complex multi-step agents are sensitive to both latency and reliability, and the infrastructure here is designed for exactly that kind of sustained, high-volume inference.

Getting Started

The official documentation covers setup from resource creation through to deployment and API calls. The Microsoft Learn guide walks through the Foundry-specific configuration steps. If you’re already using Azure AI Foundry, adding a Claude deployment follows the same pattern as any other model in the catalog.

Pricing uses Anthropic’s standard API rates, surfaced through the Azure Marketplace.