BILLING GUIDE
Every request has an accounting boundary.
How prompt and generated tokens consume prepaid credits, how reservations settle, and what happens when a stream or payment changes state.
Prompt tokens plus generated tokens
One credit pays for one prompt or generated token reported by usage.total_tokens. For example, a request with 600 prompt tokens and 200 generated tokens uses 800 credits. Tokens are model units, not characters or words; message templates also use context. Pack prices are USD and Bitcoin includes 10% additional credits.
Reservation before inference
Before sending an accepted request to inference, the gateway reserves a conservative token allowance. That keeps simultaneous requests from spending the same balance. Your available balance can therefore be lower than the balance before an in-flight request completes. Keep max_tokens within both your balance and the context envelope.
Settlement after completion
When trustworthy final usage is available, the reservation is settled to that usage and unused credits are released. The account ledger is authoritative. A third-party client's dollar estimate or model cost display is not an AblitAI bill. Requests rejected by the gateway before inference do not receive an inference reservation.
Interrupted streams
A network interruption does not prove that no tokens were generated. A timeout or disconnect after any upstream response data has been forwarded, including an initial role-only event, can be charged up to the reserved maximum when trustworthy final usage is unavailable. With no forwarded upstream response data and no trustworthy usage, the reservation is released. An explicit upstream or invalid-stream error without trustworthy usage also releases the reservation. Avoid automatic blind retries; first understand the failure and your account balance. Read the Terms for the charge boundary.
Insufficient credit and concurrency
A 402 response means the account cannot fund the request, including existing reservations or a reversal debt. A 429 may signal a rate or concurrency limit. Reduce the requested output, wait for in-flight work to settle, or inspect your available balance. Checkout may remain closed even when the balance is zero.
Payment confirmation and delayed events
A checkout page or browser redirect is not proof of payment credit. Credits require authoritative payment confirmation and validated provider events. A delayed, expired, partial, or mismatched payment may need reconciliation rather than immediate credit. Do not repeatedly create or pay invoices while a previous payment is unresolved.
Refunds, reversals and debt
Accepted refunds and payment reversals remove the corresponding purchased credits. Where already-spent credits cannot be removed from the remaining balance, reversal debt can reduce future available credit. Partial-refund handling must follow the provider-confirmed amount and cannot grant the same payment twice. Card checkout remains under review; these safeguards do not mean it is available.
Review refund terms, error examples, retained accounting records, and payment availability.