Models and pricing

• Prices are per million tokens, in US dollars. Input, cached input, and output are separate rates. A dash means the model does not publish that rate.
• The catalog row is the price. What you see for a model id is what a request to that id is metered at — there is no second, hidden rate for the same model.
• Usage comes from Responses. Token counts are returned in usage on every response, and your organization’s own usage is in the console. The public catalog does not publish per-model traffic.
• Limits are published where they exist. Context window and max output are catalog fields. Anything a model does not publish shows as “not published” rather than a guessed number.

Dashes print where a model does not publish a number. Model names link to a full page with that model’s prices, limits, data posture, and a ready-to-copy request.

{
  "model": "qwen/qwen3.8-flash",
  "input": "Summarise this incident report in three bullets."
}

Read the new model’s row first: context window, max output, and data posture can differ. Conversations that used provider-sealed state should stay on the family that produced it — see switching models safely.

cost =
    usage.input_tokens  / 1_000_000 * price.input
  + usage.cached_tokens / 1_000_000 * price.cache_read   # when published
  + usage.output_tokens / 1_000_000 * price.output

Blended averages across models are easy to misread: compare one model at a time, and always state the input/output mix you assumed.

• Training on / No training — the declared handling of API requests for this model under default routing.
• Zero retention / No ZDR — whether requests to this model use a zero-data-retention route by default. No ZDR means no such route exists for the model.
• Training-discount options — when training_opt_in is true, lower-cost options that may train on data exist for the model. They are reached only when you send x-sylphx-zdr: optional; without that header your requests never use them.
• The posture is a property of the model row, not of your key, so switching models can change it. Check the row before you move a sensitive workload.
• Platform retention for stored responses and files is separate and described in the privacy policy.