Pay as you go

By the tokens the call actually produced, at the rate the routed provider charges for that model. Input, output, cache reads, cache writes and reasoning tokens each have their own price and appear as separate lines. The cost comes back in the response body, and can also be queried afterwards by generation id.

A write is the call that stores your prompt in the cache — usually more expensive than plain input. A read is a later call that hits the cache, usually far cheaper. Some providers also price by cache lifetime (5 minutes, 1 hour); those appear as separate columns. A dash means the provider doesn't offer that component — it does not mean free.

If your client disconnects, generation upstream is cancelled but the tokens already produced are billed. Upstream cost from failed attempts during routing retries is absorbed by us — you only pay for the attempt that succeeded. Failed image and video jobs are not charged.

Add a payment method on the billing page in your console, then set a trigger threshold and a top-up amount. When your balance falls below the threshold it refills automatically. After a few consecutive failed charges it disables itself and notifies you — expired cards are normal, not an incident.

Unused credit can be refunded to your original payment method. Refunds are processed per top-up, less the payment-processing cost that top-up actually incurred. See the refund policy for the details.

Yes. Because we charge no top-up fee and absorb the processor's fixed cost ourselves, there's a minimum per top-up; the minimum for auto top-up is lower. The exact figures are shown on the top-up screen.