Changelog
Built-in AI and Decision API bill and report every answer
Built-in AI calls, Decision API questions and generated images are billed by the tokens a model reads and writes. Until now some of that work went uncounted. A model that answered before a call moved on to a backup model, a decision the model refused or answered in a shape that could not be read, and an image the provider made but sent back empty all used tokens that were neither billed nor shown in the response. Input a provider served from its prompt cache was left out too, so a call whose input came entirely from the cache was billed for its output alone. Each answer is now recorded the moment it arrives, and the usage a response reports is that same record.
How it works
- Every answer counts. When a call switches to a backup model, each model that answered is billed and included in the usage the response reports.
- Cached input at cache rates. Input a model reads from or writes to its prompt cache is billed at that model's cache rates instead of being left out.
- Refusals and unreadable decisions. A decision the model refused, or answered in a shape that could not be read, is billed and reported, since the model did the work.
- Usage you can add up. The usage in a response matches what was recorded for the call, so totals you add up from responses match your balance.
- The same through the API. Calls through the developer API are billed and reported the same way.