A Norwegian startup says it ran up a $17,600 Claude bill on Microsoft Foundry while believing its Azure credits would cover it, then spent weeks being passed between Microsoft and Anthropic support. The story, reported by The Register on October 9, is a small dispute with a general lesson: buying Claude through a cloud marketplace is not the same as buying credits-funded Azure compute, and the difference shows up on your card.
What happened, according to the report
Vegalabs AS says it received $25,060 in Azure credits through Microsoft for Startups and deployed Claude in Foundry in August, expecting its analysis workloads to draw on that balance. According to The Register, Microsoft's sponsorship rules exclude Anthropic models purchased through Azure Marketplace. Microsoft then attempted to collect $16,500 from the company's card, and the invoiced total reached $17,600 before tax.
The company says the sponsorship portal kept showing available credits while separate Marketplace charges accumulated. It deleted the deployment within an hour of discovering the charges. Its card issuer declined the payment over suspected fraud, and $21,168 of unused credits expired on September 8.
Vegalabs does not dispute the usage or the documented rule. It acknowledges it did not read the exclusions or configure a budget alert, but says the deployment interface did not make the separate billing clear. Its waiver request then bounced.
| Step | What the report says |
|---|---|
| Credits granted | $25,060 via Microsoft for Startups |
| Charges | $16,500 attempted, $17,600 invoiced before tax |
| Credits lost | $21,168 expired September 8 |
| Microsoft's reply | Marketplace is a separate billing pipeline tied to the publisher; Azure Support cannot waive, refund or adjust it; ask Anthropic |
| Anthropic's reply | Microsoft does not need its authorization to refund; ask Microsoft |
| Support quality | Four of seven replies were AI-generated, including both from Anthropic, and misdescribed the account as an "Azure Student Sponsorship" |
| Official comment | Microsoft acknowledged the query on October 6 and said nothing further; Anthropic did not respond |
Everything above is the customer's account plus messages The Register says it has seen. Neither company has confirmed the details.
Claimed vs verified
What the primary documentation confirms is the rule, not the dispute.
- Verified: Microsoft's Claude in Foundry guide lists unsupported subscription types, including "sponsored subscriptions that only use Azure credits", and adds that if you have a credit card on file, the card will be charged instead of Azure credits. It also says partner models require access to Azure Marketplace and that you accept Marketplace terms when deploying.
- Verified: The same page says you need a paid Azure subscription with an active pay-as-you-go billing method, and that student, free-trial or startup credit-only accounts are not supported.
- Reported, not confirmed: that the portal kept showing credits, that the waiver request was shuttled between companies, and the exact amounts.
- Not found: any published waiver or refund policy for Marketplace model charges. Forum threads from other Microsoft for Startups users describe similar surprise invoices, but those are anecdotes.
Why Claude on Azure bills differently
When you run a virtual machine or Azure OpenAI-style first-party service, the meter belongs to Microsoft and sponsorship credits apply. Claude models in Foundry are partner models: you subscribe through Azure Marketplace, and the publisher's terms and pricing attach to that subscription. Anthropic's Foundry documentation describes using Claude through Foundry with Anthropic's standard API pricing. Microsoft collects the money, but the product, and therefore the refund decision, sits with a third party.
That is why the two support organisations can both be partly right. Microsoft's support can see the invoice but, per its reply, cannot adjust a publisher's charge. Anthropic can set the commercial position but, per its reply, does not control Microsoft's invoicing. The customer is left in the gap. We covered a related reliance risk in Meta and Microsoft cutting internal Claude use and the naming confusion in Microsoft Foundry naming explained.
Foundry vs Anthropic direct: which channel for whom
| Question | Azure Foundry (Marketplace) | Anthropic API direct |
|---|---|---|
| Who invoices you | Microsoft, via Marketplace | Anthropic |
| Do Azure credits apply | Generally not for credit-only sponsored subscriptions | N/A |
| Where billing disputes go | Reportedly split between both companies | Anthropic |
| Good fit | Teams with a paid Azure agreement, existing procurement and Entra ID | Startups on credits, small teams, fast iteration |
| Unit prices | Anthropic's standard API pricing per its docs | List price; see Sonnet 5.5 cache pricing |
| Credits and perks | Whatever your Azure contract provides | Anthropic programs; see monthly API credits on Max and Team |
If your goal is simply to spend startup credits, Foundry's Claude is the wrong tool unless your credits are explicitly eligible. If your goal is enterprise procurement, one Microsoft invoice and committed-spend rules, Foundry may be exactly right, and the Marketplace route is the point.
What people are asking
Could a card decline have saved the company? Not really. The card issuer declined the attempted payment as suspected fraud, but the invoice remained owed, and the credits kept ticking toward their September 8 expiry. A declined card postpones collection; it does not cancel a Marketplace subscription charge.
Is this a Microsoft problem or an Anthropic problem? Both companies' documentation puts the burden of reading the rule on the customer, and both have a plausible account of why they cannot act alone. The practical failure is that nobody owns the handoff. A customer with a seven-message support trail, four of them reportedly machine-written, has no single person who can say yes.
Does this affect big companies? Less so. Enterprise buyers usually sign a Microsoft agreement in which Marketplace spend can count toward commitments, and procurement teams read exclusions before engineers deploy. The exposure is concentrated in startups, students and side projects that arrive through a credits programme and click deploy from a model catalogue.
Is the model the same on both channels? Microsoft's guide describes Claude models "Hosted on Azure" and "Hosted on Anthropic infrastructure" as deployment versions, with data-retention and regional rules that differ from the direct API. If you pick Foundry for compliance reasons, read those sections too; the billing surprise is one of several channel differences.
A pre-deploy checklist
- Read the exclusions page for your credit type. Search for "Marketplace", "partner models" and "third-party" on your sponsorship terms.
- Open the subscription's billing account and confirm which payment method a Marketplace subscription would charge. A card on file is the fallback, as the Microsoft guide states.
- Set a budget with alerts at the subscription and resource-group level. Budgets alert; they do not stop spend, so add an action group or a quota limit as well.
- Cap usage with quota, not hope. Claude deployments have default quotas by subscription tier in Microsoft's quota page, and lowering a deployment's rate limit bounds the damage of a loop or runaway agent.
- Delete test deployments the same day, and review the Marketplace subscription list separately from the usual Azure cost view.
- Keep written records of every support reply. If you must dispute, you want both companies' statements in one thread. Expect some replies to be automated; the report says most were.
- Compare against direct API pricing for your token mix before choosing a channel, including cache reads, which changed on October 8.
What this means for builders and what to watch
Agentic workloads make this kind of surprise more likely, not less. A coding or analysis agent can burn tens of millions of tokens a day, and a billing channel that sits outside your usual cost dashboards hides it. The same logic applies to any agent run unattended; if you worry about runaway actions and not just runaway cost, a guardrail layer such as AgentBeam from the explainx.ai team exists to stop agents before they take dangerous actions, though it is a different problem from invoice routing.
Three things to watch: whether Microsoft changes the Foundry deployment flow to flag credit ineligibility before you click deploy, whether either company publishes a clear Marketplace refund path, and whether the account holds up once Microsoft or Anthropic comment. The Register notes this follows a similar March case about third-party model charges not covered by Azure sponsorship credits, so the pattern is recurring rather than one-off.
For background on how Anthropic's own limits and plans shift, see our Claude usage limits timeline, and for the company's wider government and customer context, the White House incident disclosure mandate.
Bottom line
The rule was written down; the dispute shows how easy it is to miss. Before choosing a channel for Claude, decide who you want as your invoicing counterparty and who you want to argue with when something goes wrong. If the answer is "I want my credits used", go direct or confirm eligibility in writing first.
Related reading
- Meta and Microsoft cut internal Claude use
- Microsoft Foundry naming explained
- Sonnet 5.5 cache read price cut
- Anthropic monthly API credits on Max and Team plans
- Claude usage limits timeline
- Sonnet 5 permanent pricing
Details are accurate as of October 13, 2026 and rest on a single outlet's report plus Microsoft's published documentation; check current billing terms before deploying.