If you run Claude through Microsoft Foundry, your spending shows up in Azure Cost Management like any other Azure Marketplace purchase. But Claude billing has one quirk that changes how you analyze it: everything is measured in a single unit, the Claude Consumption Unit (CCU), rather than in per-model token meters. This article walks through what you will actually see in Cost Analysis, where the per-model detail lives instead, and the usage patterns that make a bill jump.
One meter, not many
Claude usage in Foundry is billed in CCUs through Azure Marketplace — a single billing dimension that replaced the previous per-model token meters. The CCU has a fixed price of $0.01, so 100 CCU equals $1.00 of fees. Your token usage is converted into CCUs using Anthropic's published per-model token rates, after any contractual discounts negotiated as an Azure Marketplace private offer. The meter is recorded hourly and invoiced monthly in arrears; there are no prepaid CCU credits or balances.
Two practical consequences follow. First, in Cost Analysis you get one Claude line per Foundry resource, not one line per model. Second, a discount does not appear as a credit on the invoice — it is applied earlier, at the token-to-CCU conversion step, so a discounted organization simply accrues fewer CCUs for the same tokens.
Filtering Cost Analysis to your Claude spend
In the Azure portal, open Cost Management > Cost Analysis at the scope you care about (subscription or resource group). Claude charges are Azure Marketplace consumption attached to the Foundry resource — the top-level object that holds billing and security configuration — not to individual model deployments inside it. Useful views:
- Group by resource to see each Foundry resource's CCU spend side by side. If different teams own different Foundry resources, this is your cost-per-team view for free.
- Group by resource group or tags to roll spend up to projects and environments (see tagging Foundry resources for cost attribution).
- Filter to marketplace charges to separate Claude and other partner models from native Azure services in the same resource group.
Because the meter is emitted hourly and marketplace charges are invoiced monthly in arrears, a heavy afternoon of traffic will not be visible in cost data instantly. Treat Cost Analysis as a daily-review tool, not a real-time one, and pair it with budget alerts for early warning.
Per-model detail lives in Foundry, not the invoice
The single CCU line answers "how much?" but not "which model, and why?". Microsoft's billing documentation points you to the Foundry portal's Monitoring tab for per-model token and request detail. That is where you break a CCU total down into, say, Opus 4.8 versus Haiku 4.5 traffic, and where you spot which deployment drove a spike. For engineering-grade analysis, also log the usage object that every API response returns — it is the ground truth for input, output, and cache token counts per request.
Spotting unexpected charges
When the CCU total looks wrong, the usual suspects are token-side, because CCUs are just converted tokens:
- Output-heavy workloads. Output tokens cost several times more than input at Anthropic's published rates (for example, $5 in / $25 out per million tokens on Opus 4.8), so a prompt change that makes responses longer moves the bill disproportionately.
- Long-context requests. The 1M-token context window carries no premium rate, but a 900k-token request still bills 900k input tokens at the standard rate.
- Model migrations. Newer models such as Sonnet 5 and Opus 4.7+ use a tokenizer that produces roughly 30% more tokens for the same text — re-baseline cost when you migrate.
- US Data Zone deployments. A US Data Zone Standard deployment applies a 1.1x multiplier on all token pricing categories, equivalent to Anthropic's
inference_geo: "us"setting. - Cache writes. Prompt cache writes bill above the base input rate (cache reads well below it), so caching that never gets read costs more, not less.
One thing that should not surprise you: on Claude Fable 5, requests refused by safety classifiers before output generation are not billed for the refused input tokens.
Where to go next
Once you can read the spend, put guardrails around it with budget alerts, and read how Azure Marketplace billing works for Foundry to understand where the CCU meter sits in your EA or MCA agreement.