Ignite runs this week (Nov 17-20), but Azure handed FinOps teams something useful before it even started. On Oct 27, Microsoft announced that cost management and tagging for Claude models in Microsoft Foundry is generally available. Claude spend now sits under the Foundry resource that runs it, so it can be tagged, grouped and broken down by model and by input and output tokens.
If you have ever tried to answer "who spent this on Claude?" from an Azure bill, that is a real gift. This post is a short lesson in how Claude billing works on Azure, what the new visibility gives you, and how to put it to work before the year-end planning cycle.
Three terms explain most of the confusion.
Claude in Foundry has been generally available since June 29, where Microsoft described CCU as a single consolidated line on your Azure bill. That is convenient for invoicing and awkward for allocation, which is the problem the October update starts to solve.
Before the update, Microsoft's message center post says, Claude charges "appeared under SaaS rather than under the Foundry resource used to deploy them." Now they follow the resource, and that unlocks four things, per Microsoft's cost management guide.
|
What you can do |
How it works |
|---|---|
|
See Claude cost per Foundry resource |
Scope Cost analysis to the resource; CCU-billed Claude consumption is associated with it |
|
Group and filter by tag |
Claude deployments and Foundry resources support tags, so team, product or cost center tags carry through |
|
Break spend down by model |
Cost records carry a model tag, so Opus, Sonnet and Haiku usage can be separated |
|
Separate input, output and cache cost |
Cost records carry a usage_type tag, which shows what kind of tokens drove the bill |
The practical value is the last two rows. Knowing that a team spent on Claude is useful. Knowing it was mostly cached input on one model, or mostly output on another, tells you what to change.
A good gift still has a few limits, and it is better to know them before you build a report on it.
Shared resources show one bucket. If three teams call Claude through one Foundry resource, Cost analysis shows one number. Giving each team its own resource avoids the problem. Where that is not practical, a Microsoft Q&A moderator suggests pulling token metrics per deployment from Azure Monitor, weighting them by each model's rate, and splitting the billed total in proportion. Treat that as an estimate, not an invoice line.
The docs describe project-level attribution for Marketplace models as not yet supported and publish no timeline. That makes it a good thing to listen for this week.
The broader lesson is that AI spend is arriving on cloud bills faster than allocation models adapt. The teams that can answer "who spent this" are the ones that settled on an allocation rule before anyone asked. That is where Finout's virtual tags help: they let you allocate spend with rules on top of billing data, instead of depending on every resource being tagged perfectly at creation.