prompt_tokens meant two things. The family and relay paths recorded the
part of the prompt read fresh, the table path and the providers the
whole prompt, so a cached enso-auto row read cache_read > prompt. No
path recorded a cache write: Anthropic's cache_creation_input_tokens
was folded into the fresh read and billed at the input rate, though a
write is priced above it, and OpenRouter's cache_write_tokens was not
read at all. cache_write_tokens was 0 on every row.
A record's prompt is now every input token, with the cache reads and
writes parts of it, whatever shape the upstream reported: OpenAI's
cached_tokens inside prompt_tokens, OpenRouter's cache_write_tokens
beside them, Anthropic's two counts beside input_tokens. Pricing bills
the input rate on the part read fresh and each cache part at its own
rate: a family SKU's cache_read and cache_write (a resold model's
input_cache_read and input_cache_write plus the vendor's fee), else the
input rate. The Anthropic provider and the OpenAI adapter over
Anthropic now read both cache counts, and the providers' settles pass
them.