Claude on Microsoft Foundry sits inside your Azure subscription, so when something breaks, your path to help is the Azure support system — governed by whichever Azure support plan your organization pays for. Microsoft's plans range from free basic support (billing and subscription help only) through paid tiers historically named along the lines of Developer, Standard, Professional Direct, and Premier/Unified, each adding faster initial response targets, around-the-clock coverage for severe incidents, and more proactive services. The specific response times, prices, and tier names change over time and are not something this guide will pin down — check Microsoft's current support plan documentation and your agreement for the numbers that would apply to you. What this article can do is map those generic tiers onto the realities of running Claude on Foundry.
Which tier fits which Foundry workload
The honest sizing question is not "how big is our Claude bill" but "what happens to the business while a Claude-dependent flow is down."
| Situation | Sensible floor |
|---|---|
| Prototypes and internal experiments | A developer-grade plan — business-hours help is enough when nothing external depends on the workload |
| Internal production tools (assistants, document processing) | A standard-grade plan with 24/7 coverage for high-severity cases |
| Customer-facing or revenue-bearing flows | The higher tiers with the fastest critical-severity response and an account relationship — and a tested failover design, since support response time is not outage recovery time |
The two-vendor reality behind a Foundry ticket
Claude models in Foundry are third-party Azure Marketplace offerings: Anthropic is the seller and operator of the models (for both the Hosted-on-Azure and Hosted-on-Anthropic options), while Microsoft provides the platform, and Anthropic's docs describe both hosting options as Anthropic-operated services supported through the Azure relationship. In practice you open incidents through Azure support, and the two companies trace the request across their systems. That's why every Foundry response carries two correlation headers: request-id (Anthropic side) and apim-request-id (Azure side). Anthropic's documentation says to provide both to support — so make sure your application logs both on every failure, or your first support exchange will be spent reproducing the problem instead of diagnosing it.
What's a support ticket, and what isn't
A useful amount of "Foundry isn't working" traffic is not an incident at all, and knowing the difference saves your severity credibility for real ones.
Quota and rate limits. 429 responses under load usually mean you've hit your subscription's requests-per-minute or input-tokens-per-minute quota — and some defaults are surprising, such as claude-fable-5 starting at zero on pay-as-you-go subscriptions. Increases go through Microsoft's quota increase request form, a separate process from support incidents.
Access errors. 403s typically mean a missing role assignment (such as Cognitive Services User) on the Foundry resource; "Region not available" means the deployment targeted an unsupported region. Both are configuration fixes.
Refusals. A response with stop_reason: "refusal" is the model's safety system working as documented, not a platform fault.
Genuine incidents — elevated error rates, latency regressions, endpoint unavailability — are where the plan's response targets matter. When you file one, include the deployment name, region, both request IDs, timestamps in UTC, and whether the deployment is Hosted on Azure or Hosted on Anthropic infrastructure, since the two run on different infrastructure paths.
Procurement notes
Support plans are subscription-level Azure decisions, usually owned by the platform or cloud team rather than the AI team — if your Claude workload is the first business-critical thing in a subscription that only ever ran experiments, the plan may lag the workload's importance. Also keep expectations precise: a support tier is a response-time commitment for tickets, not an availability guarantee for the service. For what availability commitments do and don't cover, and how to architect around them, see disaster recovery planning for Foundry-dependent applications.
Where to go next
Read Foundry error codes to triage before you file, and Foundry quota increases for the limits path. The production checklist covers the rest of go-live readiness.