Dual-running — sending live traffic to two Claude platforms at once, whether for failover, migration, or negotiation leverage — is a sound engineering practice with an annoying accounting side effect. The four third-party platforms split into two billing families, and the numbers they put on your invoice are not directly comparable line for line.
The two billing families
Per-token, billed by the cloud provider. Amazon Bedrock and Google Vertex AI meter your usage in tokens and bill you directly on your AWS or Google Cloud invoice, at per-million-token rates published on each provider's pricing page. What you see is close to what the Claude API itself would show: an input-token rate and an output-token rate per model, plus cache-write and cache-read rates.
CCU, metered through the marketplace. Claude Platform on AWS and Claude in Microsoft Foundry bill in Claude Consumption Units (CCUs) at $0.01 per CCU — 100 CCU equals $1.00 of fees at the standard per-model rates, after any discounts. Usage is metered hourly to the respective marketplace (AWS Marketplace or Azure Marketplace) and invoiced monthly in arrears. CCUs are not prepaid credits; there is no balance to top up.
| Platform | Billing unit | Invoice lands in | Per-model detail |
|---|---|---|---|
| Amazon Bedrock | Tokens | AWS bill | Per-model token line items |
| Google Vertex AI | Tokens | Google Cloud bill | Per-model token line items |
| Claude Platform on AWS | CCU ($0.01 each) | AWS Marketplace | Claude Console usage/cost pages |
| Microsoft Foundry | CCU ($0.01 each) | Azure Marketplace | Foundry portal Monitoring tab |
The good news: because a CCU is a fixed $0.01, converting a CCU line back to dollars is trivial. On Foundry, token usage is converted to CCU using Anthropic's published per-model token rates, with any negotiated discount (delivered as an Azure Marketplace private offer) applied at the token-to-CCU conversion step — so the CCU quantity on the invoice already reflects your discount.
Build the model on tokens, not invoices
The one thing all four platforms share is the response format: every Messages API response carries the standard usage object with input, output, cache-write, and cache-read token counts, consistent across platforms. That is your common denominator. A workable unified spend model looks like this:
1. Log the usage object on every request, tagged with platform, model, and workload. This gives you a platform-neutral record of consumption in the only unit that means the same thing everywhere: tokens.
2. Price those tokens at Anthropic's list rates (for example, Opus 4.8 at $5 input / $25 output per million tokens; Haiku 4.5 at $1 / $5). Cloud marketplace list prices match Anthropic's first-party list prices, so this "shadow rate card" is a fair baseline for every platform before adjustments.
3. Apply per-platform adjustments — this is where dual-run comparisons usually go wrong. The adjustments that differ by platform include:
inference_geo: "us" and applies a 1.1x multiplier. And Sonnet 5 has introductory pricing ($2/$10) only through August 31, 2026 — a dual-run comparison built in July will drift in September.4. Reconcile monthly against the real invoices. Your token-based estimate should land close to the AWS/Google token lines and, after dividing CCU quantity by 100, close to the marketplace CCU lines. Persistent gaps usually trace to caching (cache reads bill at 0.1x input), batch discounts, or one of the multipliers above.
Where to find each platform's own numbers
For the CCU platforms: Foundry shows one CCU line in Azure Cost Management, with per-model token and request detail in the Foundry portal's Monitoring tab. Claude Platform on AWS bills through AWS Marketplace, and the Claude Console's usage and cost pages show organization-level detail — note that Anthropic's Usage and Cost API endpoints are not available on that platform, so plan on console review or your own request-level logging rather than API-driven pulls. For the per-token platforms, Bedrock costs appear in standard AWS cost tooling and Vertex costs in Google Cloud billing, which is exactly why request-level logging with the usage object matters: it is the only view that spans all four.
Where to go next
For the mechanics of each invoice, see Claude Platform on AWS billing, Foundry billing mechanics, and Vertex billing mechanics. If committed-spend programs are steering your platform choice, read CCU billing, MACC eligibility, and platform selection next.