API credits vs token packs
How prepaid AI API credits and token packs differ: what each request deducts, how cached tokens are counted, and which CloudService route each key uses.
Credits and tokens are different meters. A credit key deducts a customer-facing monetary balance; a token key deducts a fixed token allowance. Neither should be confused with the provider cost paid by the gateway.
Keep three ledgers separate
A transparent AI gateway records three independent views of every completed request:
- Raw token accounting: input, cached input, uncached input, output, and total provider-counted tokens.
- Customer billing: the published-price credit or token allowance deducted from the key.
- Provider cost: the gateway's internal upstream cost, which is not exposed as the customer's bill.
Combining those values into one number makes dashboards, limits, and margins disagree.
Credit keys deduct published-price usage
A credit key has a USD-denominated internal usage balance. Input and output are priced at the customer-facing rate published for the selected model. Provider cache discounts do not silently change that public rate.
OpenAI-compatible clients use https://api.cloudservice.services/v1.
Native Claude clients use https://api.cloudservice.services.
This mode is useful when one balance should cover supported models across more than one family and the user prefers monetary usage reporting.
Token keys deduct a provider-scoped allowance
A token key starts with a fixed allowance. Each completed request deducts the exact provider-counted input and output token totals available from the response. The key remains tied to the family or exact-model entitlement selected at purchase.
OpenAI-compatible clients use https://api.cloudservice.services/token/v1.
Native Claude clients use https://api.cloudservice.services/token.
A Claude token key does not become an OpenAI token key because the URL looks compatible. The authenticated model catalog is the source of truth.
Cached input still belongs in raw token accounting
Suppose a provider reports 10 million input tokens, of which 9 million were cache reads, plus 2 million output tokens. The raw ledger stores 1 million uncached input, 9 million cached input, and 2 million output. The provider-total count is 12 million.
Cache may change the provider's internal cost, but it does not erase tokens from the request. If a provider does not report a cache count, the system should mark cache data unavailable rather than inventing a number.
Which key should you choose?
| Choose | When it fits | Primary dashboard unit |
|---|---|---|
| API credit | You want one monetary balance for supported routes and models. | Credits used and remaining |
| Token pack | You want a fixed provider-family or exact-model allowance. | Tokens used and remaining |
Decide the target client and model family before checkout. The CloudService guided setup and documentation then show the matching base URL without requiring the user to memorize route rules.