Claude Platform on AWS is the unusual member of the 3P family: the models run on Anthropic-managed infrastructure with same-day API parity (with documented exceptions), while AWS provides authentication, IAM-based access control — and billing. Signing up on the AWS Console's Claude Platform on AWS service page has AWS handle the Marketplace subscription and provisions a brand-new Anthropic organization tied to your AWS account. From then on, usage flows to your AWS invoice, not to an Anthropic one.
CCUs: the billing unit
Charges are denominated in Claude Consumption Units (CCUs), metered hourly to AWS Marketplace and invoiced monthly in arrears. CCUs are not prepaid credits — there is no balance to top up or burn down; this is pure postpaid consumption. Anthropic's pricing page pins the unit: $0.01 per CCU, so 100 CCU equals $1.00 of fees at standard per-model rates after discounts. Your token usage at Anthropic's published per-model prices (Opus 4.8 at $5/$25 per million input/output tokens, Sonnet 5 at $3/$15 standard with $2/$10 introductory pricing through August 31, 2026, Haiku 4.5 at $1/$5, Fable 5 at $10/$50) is converted into CCU quantity, and CCU quantity times $0.01 is the dollar line on the bill.
Standard pricing modifiers apply upstream of the conversion: prompt cache writes at 1.25x or 2x base input and cache reads at 0.1x, batch processing at 50% of standard rates, and the 1.1x multiplier for inference_geo: "us" (US-only inference) on supported models. One documented exception: fast mode is not available on Claude Platform on AWS at all.
Where the charges appear
Because the meter runs through AWS Marketplace, the charges appear in your AWS cost tooling alongside native services, attributed to the AWS account that subscribed — which makes account-per-team structures inherit attribution automatically. Within an account, the platform's own unit of organization is the workspace: usage, quotas, cost, files, batches, and Skills all roll up per workspace, and each workspace is bound to a single AWS region. Workspaces are therefore your natural chargeback dimension — one per team, product, or environment — enforced by IAM policies on workspace ARNs.
The controls you won't find
Several cost-management features of the first-party Claude API are documented as unavailable here, and your FinOps design has to route around them:
| First-party feature | On Claude Platform on AWS |
|---|---|
| Spend limits (monthly caps) | Not available — rely on AWS billing controls |
| Usage & Cost Admin API | Not available (only Admin API workspace endpoints exist) |
| Automatic usage-tier movement | Organizations stay on the Start tier; contact Anthropic to raise limits |
| Claude Code Analytics API | Not available; Claude Code usage appears in the general usage view |
What you do get is the Claude Console, federated through AWS IAM (the aws-external-anthropic:AssumeConsole action): its Usage, Cost, Limits, and Workspaces pages read organization metadata directly from Anthropic, so humans can inspect spend per workspace even though there is no programmatic cost endpoint. For automated pipelines, capture the usage object from every API response in your own logs — the approach in logging token counts for internal FinOps — and multiply by list prices.
Practical setup checklist
First, decide account and workspace boundaries before launch — they are your only native attribution dimensions, and retrofitting is painful. Second, since there are no Anthropic-side spend caps, configure AWS-side budget alerting on the Marketplace charges and treat rate limits (managed by Anthropic, Start tier by default) as your burn-rate ceiling. Third, note that CloudTrail logs inference calls as data events (paid, explicit opt-in), while workspace management actions are management events logged by default — relevant if you plan to derive usage attribution from CloudTrail. Finally, reconcile early: run a known workload, compare Console cost pages, your token logs, and the Marketplace line, and confirm the CCU arithmetic before finance sees the first real invoice.
Where to go next
Contrast this with Bedrock's billing mechanics on the same cloud, then design attribution with a chargeback model or anomaly detection for Claude spend.