Sign in

Billing and usage

Each request costs the provider’s list price for the tokens or units it used. Nothing is added per request.

How a request is billed

  1. Before it runs, AI Runtime holds the most the request could cost from your balance. If the balance can’t cover it, the request is refused with quota_model_budget_exceeded and costs nothing.
  2. When it finishes, the provider’s reported usage is priced at the list rate and the hold is settled to that amount. It is never more than the hold.
  3. Cached input (usage.prompt_tokens_details.cached_tokens) is billed at the cached-input rate. Reasoning tokens are output tokens, billed at the output rate.

The cost on every response

Chat completions include airuntime.cost_usd_micros, the request’s cost in millionths of a US dollar (2150 is $0.00215). Speech responses carry it in the x-airuntime-cost-usd-micros header instead. Streams include it in the usage chunk when you set stream_options.include_usage.

Failed requests

  • A request that fails outright, refused by the API or by the provider, releases its hold: you pay nothing.
  • If the outcome is uncertain (the provider may have run it, or a stream stopped partway), the hold stays until it is reconciled with the provider’s usage. A stream that stops without a usage report is not settled as zero.

Peak hours

DeepSeek peak is Mon–Fri 01:00–04:00 and 06:00–10:00 UTC and costs double. Every other hour, and all weekend, is off-peak. The rate in force when a request runs is the rate you pay; model pages show both.

Your usage

Your Dashboard shows your balance, spend by day and by model, your cache hit rate and your latest calls, as soon as the provider reports them. The 5% fee (minimum $0.79) is charged once, when you add credit; see pricing.

Get a key, then make your first call.

From $5, plus a 5% top-up fee (minimum $0.79). No subscription; paid credit never expires.

Add credit