6 min read · updated 2026-08-31 · reviewed

Prepaid AI API pricing: tokens, reservations and limits

Prepaid access is easiest to reason about when the public price, temporary reservation and terminal usage record agree. Here is the Model.sale billing flow in plain language.

Editorial note

Written by Model.sale Engineering and reviewed against first-party documentation on 2026-08-31. Comparisons use the billing rules visible on the linked sources and do not claim universal model equivalence.

Price per million charged tokens

Each published model has a USD blended rate per one million charged tokens. The usage ledger keeps input, cached, output and reasoning fields separately when the protocol reports them; the public rate is not an official split input/output tariff.

This distinction matters: an official page may show one input rate and a much larger output rate. A blended Model.sale number must be applied to the same full workload, not compared only with the official input column.

Temporary reservation

Before dispatch, a safety amount is reserved so concurrent requests cannot overrun the wallet. When terminal usage arrives, the final charge is calculated from measured tokens and the unused reservation is released.

The reservation shown in the UI is therefore a temporary hold, not a minimum final charge. If terminal usage is missing, the request remains pending for reconciliation instead of inventing a token total.

A worked example

Suppose a job uses 100,000 input tokens and 20,000 output/reasoning tokens. At a blended rate of $0.09 per million charged tokens, the estimate is 120,000 ÷ 1,000,000 × $0.09 = $0.0108. The actual ledger follows terminal usage, which may differ from the estimate.

For an official split tariff, multiply input and output separately, then add cache writes, cache reads, tools and payment fees if the provider charges them. Use the workload calculator to keep the units consistent.

Limits are part of cost control

RPM, TPM, concurrency and spend caps protect both your balance and the service. Create separate keys when an application needs a different operating envelope.

A single-request cap prevents a mistaken output limit from reserving too much balance. A daily or monthly cap limits the effect of leaked credentials, runaway retries or an agent loop.

How to compare services

Record the exact model ID, token mix, cache policy, retry behavior, payment fee and check date. Then compare the resulting total together with availability and latency. Avoid universal percentage-saving claims when the billing units differ.

Sources and further reading

Checked on 2026-08-31. These links support the factual comparisons in this article; pricing and terms may change.

Use the numbers on your own workload

Open the cost calculator, choose a live model and enter a representative input/output mix before deciding.

Open cost calculator
← All guides