Text requests are billed per token. Input and output are priced separately, and cached input is priced separately again — usually far below the uncached rate. Some models, mostly image and video, charge a flat price per call regardless of length.
What gets counted
| Component | Billed as |
|---|---|
Input tokens | Your prompt, at the model's input rate |
Output tokens | The reply, usually a multiple of the input rate |
Cache read | Input served from cache, at a steep discount |
Cache write | Input written into the cache, at a small premium |
Every request lands in your usage log with its token counts and the rates that produced the charge, so any line item can be reconstructed rather than taken on trust.
Which account pays
A personal key draws on your own balance. An organization key draws on that organization's wallet, and if the member holding it has an allocated budget, it draws on the budget first. Owners can reclaim whatever is left unspent.