一个页面的成本
What a page costs
我们对外报出的数字：生成一个页面、一次修改、一张照片描述、一次提问回答分别花多少。这些数字从何而来，提示缓存改变了什么，以及为什么它只是估算而非账单。
成本
Clients pay for model usage, so the system has to be able to say what a page cost. This note gives the figures we quote and explains where they come from.
The figures
| Call | Model usage |
|---|---|
| a page from first paste to published | usually well under one US dollar of model usage |
| a revision turn | a few cents, because the family's prompt is cached and only the new material is paid at full price |
| a photo description from the vision model | a fraction of a cent |
| an answer from the question box | about half a cent, on the smallest model with a fixed-size prompt |
| an embedding for a changed block | negligible |
Why the composer is the expensive part
The composer runs on the most capable model, because layout judgement and faithful restructuring of a long source are where cheaper models fall short. It returns the whole page every turn, which is simple and safe but means output tokens scale with page length. A two thousand word report is a few thousand output tokens. At current list prices that is the bulk of the dollar.
Where caching helps
The system prompt for a family, the rules, the block documentation and the voice, is identical from turn to turn and is cached. On a revision turn the model reads the cached prompt at a tenth of the input price, and the new material is the only full-price input. This is why a client can ask for five rounds of changes without the cost of five pages.
How the site keeps count
Every model call records its model and token usage in an audit log. A price table in the server, per million tokens for input, output, cache read and cache write, turns those into an estimated dollar figure. Unknown models are priced at the top tier so the estimate errs high. The admin usage page shows totals for the last thirty days by kind of call, today's question-box count and spend against the daily limits, and the estimate for each call.
An estimate, not an invoice
List prices change and the table is maintained by hand. The figure is for budgeting and for the conversation with a client, not for billing to the cent. For billing, the vendors' own dashboards are the source, and the monthly caps set there are the backstop that holds if the site's arithmetic is wrong.
What this means for a client
A client publishing a few pages a week spends a few dollars a month on composition. The infrastructure, a small Cloud Run service and a Firestore database at this scale, is of the same order. The question box, capped at two dollars a day on this site, cannot exceed sixty a month even if every cap is reached every day, and in practice it costs cents.