Track inference costs, limits and receipts
Distinguish reservations, measured usage and charges. Read RN receipts, interpret pending states and avoid repeating paid calls.
Rota Nacional · Published · Reviewed
Tokens do not describe the whole bill
A request may reserve funds before the answer and finish with lower consumption. The reservation protects the authorized ceiling. Settlement uses the contract and measurable usage. Reserved and charged amounts therefore appear as separate information.
Provider cost and RN service cost are also different. An owned key and a supplied key may follow different contracts. Check the tariff that applies to your case. A missing value does not mean zero cost.
Recover state before repeating work
When the gateway creates a reservation, X-RN-Inference-Receipt identifies its receipt. The /v1/inference/receipts route reads the payer's authorized records. Receipts are also available in inference management.
An interrupted connection does not prove that the provider did not execute the request. Unknown usage can remain pending reconciliation. Check the receipt and asynchronous resource state before submitting the same work again. RN does not automatically repeat a submitted native request.
Set limits at more than one level
An output limit bounds the requested answer size. Organization quotas control requests and tokens over a period. Balance and contract control financial authorization. These controls complement one another; none replaces the others.
Files and reads are not equivalent to a new inference. Batches need a contract that accounts for later processing; a token tariff at creation does not measure all batch work. When investigating an error, retain request and receipt identifiers without putting prompts or keys in public logs.
Frequently asked questions
Are reserved and charged amounts the same?
Not necessarily. A reservation covers the authorized ceiling. Settlement uses the contract and measured usage; the difference can be returned.
Does a connection failure mean I should repeat the call?
Not by itself. The provider may have received the request. Check its receipt and work state before trying again.
Keep reading
An AI gateway that preserves your application's model
Centralize AI connections in RN without converting provider request bodies. Understand native endpoints, authorized pools and migration limits.
Manage multiple AI keys with organization isolation
Organize provider connections by company. Learn how pause, weights, consent and file ownership keep customer usage separate.
AI credit by Pix: continue after your trial
Understand RN credit purchase, organization balance and usage charges. Payment opens over the task and preserves your work context.