Pricing
How model prices map to what you're billed.
How prices work
The published price is the price charged. Prices reported by GET /v1/models (and shown in the catalog) are quoted in USD per 1M tokens, separately for prompt and completion tokens — those are the exact rates you're billed at, for the bare model IDs the catalog lists. A provider-prefixed ID (e.g. <provider>/claude-haiku-4-5-20251001) is still accepted, but bills at that upstream's real rate instead of the published one.
A model offered at several context windows is listed once per window, under a context-suffixed ID — hypothetically, some-model-200k and some-model-1m for a model offered at both 200k and 1M tokens — each priced separately.
The Auto Mode IDs (auto*) publish their price band's cap. That cap is what's reserved and the most you can be billed; where the upstream reports the real cost of the model it picked, you're charged that instead.
What you're charged
- Charges are computed from actual token usage.
- Balance is reserved before the call based on
max_tokens, then reconciled to the real cost. See Credits & Billing. - On an upstream error, the reservation is released and no charge is made.