Claude is now generally available in Microsoft Foundry — hosted natively on Azure, billed through Azure, and governed by the identity and access policies enterprises already operate. The announcement, which came in two steps on June 29 and July 1, removes the last major institutional barrier to enterprise Claude deployment: the need to sign a separate Anthropic commercial contract and manage a second vendor relationship. Enterprise teams with existing Microsoft Azure Consumption Commitments can now draw down that pre-committed spend against Claude usage — a financial path that did not exist when procurement required going directly to Anthropic.

The launch brings three Claude models into Azure-hosted production in rapid succession. Claude Opus 4.8 and Claude Haiku 4.5 reached general availability in Foundry on June 29, with Claude Sonnet 5 — Anthropic’s newest Sonnet model and now the default across Claude’s consumer and team plans — following on July 1.

What the GA Actually Changes

Most enterprise AI projects do not stall because of model quality. They stall, as Azure Product Lead Steve Sweetman wrote in the June 29 Azure blog announcement, because of “everything around the model: procurement, governance, networking, and data.” Those are precisely the dimensions this launch addresses.

Before this week, an enterprise developer wanting to deploy Claude through Azure could already do so — Anthropic’s API was always reachable — but the billing, identity management, and security review sat outside Azure’s governance envelope. A central IT team following standard vendor approval processes had to treat Anthropic as a distinct vendor with its own security assessment, data processor agreement, and invoice. That friction kept many organizations in a pilot-indefinitely state: the model capabilities were there, but the path to production was not.

The June 29 and July 1 releases collapse that friction. Teams authenticate with Microsoft Entra ID, apply Azure role-based access controls, and receive a single consolidated Azure invoice with per-model detail. Claude usage is billed in Claude Consumption Units (CCU) as a single line on the Azure Marketplace bill. For eligible customers with a Microsoft Enterprise Agreement, that usage draws down existing Microsoft Azure Consumption Commitment credits — converting a previously off-ledger AI tool into a spend category their finance teams already manage.

How Azure Governance Solves the Shadow AI Problem

There is a structural problem in enterprise AI adoption that procurement simplification quietly solves: shadow AI. When central IT cannot approve an external AI tool quickly enough — because the security review, data processor agreement, and vendor vetting take months — product teams and developers use it anyway through personal accounts or unofficial channels. The result is a governance gap: usage the organization cannot audit, data it cannot account for, and cost it cannot attribute.

Claude’s integration into Azure’s existing governance stack directly addresses this pattern. IT organizations that have already approved Azure as a platform can sanction Claude use within their existing framework, without a new security review, because the relevant controls — Entra ID authentication, role-based access controls, private networking, and data residency configuration — are all managed through the same governance interface they already operate.

For high-sensitivity workloads, zero data retention is available: prompts and completions are not retained by Anthropic after an API call completes. Customers can choose between Global and US data zones for data residency requirements. Anthropic operates the inference and is the data processor and SLA provider.

One operational boundary worth noting: Claude models on Azure-hosted Foundry currently deploy in East US 2 and Sweden Central only. Startup accounts, free trials, and Azure subscriptions without a pay-as-you-go billing method are excluded — the Azure Marketplace requirement means organizations need an active paid Azure subscription.

The Hardware Layer: What NVIDIA GB300 Blackwell Ultra Means for Inference

Claude runs on NVIDIA GB300 NVL72 systems connected by Quantum-X800 InfiniBand networking — a combination that matters not for its marketing weight but for what it enables architecturally.

The GB300 NVL72 is a rack-scale system: 72 Blackwell Ultra GPUs and 36 Arm-based Grace CPUs liquid-cooled into a single coherent compute domain, with all 72 GPUs connected by fifth-generation NVLink. A single rack delivers 1.1 exaFLOPS of FP4 compute and 1.5 times the AI performance of the prior Blackwell generation. The architecture is deliberately inference-weighted: Blackwell Ultra adds 50 percent more NVFP4 compute and doubles attention-layer acceleration versus standard Blackwell — the specific operations that dominate transformer-based model inference. This is why a single GB300 rack delivers up to 50 times the AI factory output performance of a comparable Hopper-generation system.

The Quantum-X800 InfiniBand fabric is what makes the 72-GPU configuration viable at low latency. InfiniBand uses Remote Direct Memory Access to transfer data directly between the memory spaces of networked systems, bypassing the operating system kernel entirely. This reduces interconnect latency to 3 to 5 microseconds, compared to 20 to 80 microseconds for standard Ethernet — a gap that matters enormously when dozens of GPUs must stay synchronized during inference. Without high-speed InfiniBand, the network becomes the bottleneck that prevents GPU compute from being fully utilized.

For enterprise teams deploying Claude for high-volume agentic workloads — the use case Foundry Agent Service is built around — this hardware architecture directly determines throughput capacity and per-token cost at scale.

Model Router: How Foundry Cuts Up to Half of Inference Costs

Microsoft’s model router capability, available within Foundry, addresses one of the practical cost problems in enterprise AI deployment: most production workloads are mixed-complexity, but most organizations pay premium-model prices for every request regardless of what that request actually requires.

The model router is itself a trained machine-learning model — not a rules engine — that analyzes each incoming prompt, including the system message, user message, tool definitions, and conversation history, and routes it in real time to the most appropriate model in a configured pool. Simple prompts route to faster, cheaper models; complex prompts requiring extended reasoning route to premium models. The routing decision adds negligible overhead — a fraction of the underlying inference time. The result, according to Microsoft, is up to 50 percent cost savings while improving user satisfaction, because simpler requests also return faster on lighter models.

This architecture reflects a genuine engineering insight: the bottleneck in enterprise AI economics is not the existence of capable models but the absence of intelligent dispatch. Foundry Control Plane continuously evaluates agent responses against customer-defined rules and can block outputs that violate those rules before they reach users — adding a governance layer that manual model management cannot replicate at scale.

Claude Sonnet 5 in Foundry: What Changed on July 1

Claude Sonnet 5, which Anthropic released on June 30 as the new default model across its consumer and developer plans, reached GA in Microsoft Foundry on July 1. Built as an upgrade to Sonnet 4.6, Sonnet 5 is positioned as a meaningful step up in agentic performance: stronger tool use, better multi-step coherence, and improved reliability on coding and document workflows that previously required Opus-class models.

For enterprises, Sonnet 5 arrives in Foundry at a promotional price of $2 per million input tokens and $10 per million output tokens — in effect through August 31, 2026, after which standard pricing of $3 per million input tokens and $15 per million output tokens takes effect. The promotional pricing uses a new tokenizer that encodes roughly 30 percent more tokens for the same text compared to Sonnet 4.6, so the apparent per-token discount is approximately cost-neutral with the prior model at standard rates. Teams evaluating migration should benchmark against their own prompt distributions rather than assuming a flat savings.

The competitive context is direct. OpenAI’s models have been available in Azure OpenAI Service for years, benefiting from the same SSO, policy, and billing integration Claude now has. Anthropic’s GA closes that governance-parity gap on the procurement and identity dimensions enterprise buyers prioritize.

Production Teams Already Running Claude in Foundry

Early production users illustrate the scale at which the Foundry integration operates. Momentic, an AI testing platform, reported through its co-founder Jeff An that running Claude’s Opus models on Microsoft Foundry lets the company serve millions of tokens per minute with the reliability its customers depend on.

Matt Huang, Founding Product Lead at Everstar, described the practical scale of what Claude enables in regulated technical environments. The integration allowed the nuclear technology company to compress a safety analysis that would have taken 200 human days into a single day. “Between Anthropic and Azure, we get the best capabilities in the world and we get the best security in the world,” Huang said. “And that’s exactly what nuclear needs.”

The November 2025 strategic partnership between Microsoft, NVIDIA, and Anthropic provided the foundation for this week’s launches — the GB300 hardware deployment and NVIDIA’s involvement in the Foundry Agent Service architecture both trace to that earlier agreement.

Microsoft IQ, Foundry Agent Service, and What Microsoft Gets From This

The enterprise framing of this launch extends beyond model access. Microsoft IQ gives agents access to live enterprise context — connecting Claude’s reasoning to data already inside Microsoft’s ecosystem — which, according to Microsoft, radically improves value per token. An agent optimizer within Foundry Agent Service tunes the prompts that define agent behavior, improving performance regardless of which underlying model is in use.

For Microsoft, the architecture reflects a deliberate strategic position: own the governance layer, the billing relationship, and the identity infrastructure, then remain model-agnostic in the frontier model tier. Foundry is currently used by developers at more than 80,000 enterprises, including 80 percent of Fortune 500 companies. Adding Claude — and positioning Foundry as the single platform offering both Claude and GPT-class frontier models — makes the platform itself, not any individual model, the durable enterprise relationship.

Frequently Asked Questions

Does deploying Claude in Microsoft Foundry require a separate contract with Anthropic?

No. Enterprises can access Claude through an existing paid Azure account, with billing, identity management, and governance handled through Azure. For eligible customers with a Microsoft Enterprise Agreement, Claude usage draws down existing Microsoft Azure Consumption Commitment credits. Free trial accounts, startup sponsored subscriptions, and accounts without active pay-as-you-go billing are not eligible.

What is Claude Sonnet 5 and what does it cost in Microsoft Foundry?

Claude Sonnet 5 is Anthropic’s newest Sonnet-class model, generally available in Microsoft Foundry as of July 1, 2026. It is designed for stronger agentic performance than Sonnet 4.6, with improvements in tool use, multi-step task completion, and coding workflows. Promotional pricing through August 31, 2026, is $2 per million input tokens and $10 per million output tokens, after which standard pricing of $3 and $15 per million tokens applies. Sonnet 5 uses a new tokenizer that encodes roughly 30 percent more tokens per equivalent text compared to Sonnet 4.6.

What does NVIDIA GB300 Blackwell Ultra hardware mean for Claude inference performance?

The GB300 NVL72 is a rack-scale system with 72 Blackwell Ultra GPUs connected by fifth-generation NVLink and Quantum-X800 InfiniBand networking. Blackwell Ultra adds 50 percent more low-precision compute and doubles attention-layer acceleration versus prior Blackwell hardware — the operations most critical for large language model inference. The InfiniBand networking uses Remote Direct Memory Access to cut interconnect latency to 3 to 5 microseconds, preventing network bandwidth from becoming the bottleneck on high-throughput multi-GPU inference.

Is Claude in Azure the same as Claude hosted directly on Anthropic’s infrastructure?

Not entirely. The Azure-hosted option currently supports Claude Opus 4.8, Claude Haiku 4.5, and Claude Sonnet 5 in Global Standard deployments in East US 2 and Sweden Central. A hosted-on-Anthropic option remains available for models or API features not yet available on the Azure-native path. Microsoft and Anthropic have stated their goal is feature and model parity between the two options over time.