Model rates
How AUTO and provider model usage is priced per million tokens.
Model usage bills at the published per-million-token rate of the model that served each request.
Where rates come from
Input and output rates come from the published models.dev model catalog. That catalog is the versioned source: every entry carries the date it was last updated, and the platform reads the current catalog instead of a hand-maintained list. The machine-readable catalog is at models.dev/api.json.
Looking up a rate
A catalog entry prices one model at one provider. It reads as a model key and its rates, in US dollars per million tokens:
["togetherai/deepseek-ai/DeepSeek-V4.1-Flash", 0.3, 1.2, 0.006]togetherai/deepseek-ai/DeepSeek-V4.1-Flash- the provider and model key.0.3- the input rate.1.2- the output rate.0.006- the cache-read rate, listed only when the catalog publishes one.
In the catalog's own JSON the same entry reads cost.input, cost.output, and cost.cache_read. The catalog dated that entry 2026-09-10.
Rates change when a provider changes them, and the same model id can carry a different rate at a different provider, so look up your exact provider and model key in the current catalog.
Cached prompt tokens bill once, at the cache-read rate when the catalog publishes one and at the input rate when it does not. They are not also billed as input.
AUTO and AUTO-MAX
AUTO and AUTO-MAX route each request to an underlying model, and usage bills at the rates of the model that served it, so both tiers cost what the serving model costs.
A request that does not report the model that served it prices from the tier's own catalog entry. A model key the catalog no longer lists - for example a model an earlier release routed - is handled by the platform's pricing resolver, which prices it from a related catalog entry when one exists.
Other provider models
Other provider models use their published catalog rate when a match exists.
Fallback
If no catalog match exists, a small platform minimum charge applies per usage event. An event that reports no billable tokens is not charged.