AI Usage Cost: What You Should Measure
What should you measure so the AI bill is no surprise? Tokens, caching, retries and cost per business outcome, as a measurement plan that invents no prices.

A team starts trying an AI-assisted feature. The first month's bill is small. By the third month user numbers have grown, prompts have lengthened, and the bill has reached a figure nobody calculated. To control cost you first need to know what you are measuring.
What the bill is based on
Model providers usually price by token. Anthropic's pricing page lists models in separate columns: base input tokens, cache writes (for two different durations), cache hits and output tokens; prices are per million tokens and vary by model. The page also points elsewhere for current pricing, so these numbers change over time. We deliberately give no prices here; open the page yourself: Anthropic pricing page (read on 9 November 2026).
What matters: a single request creates cost in several separate items, and their unit prices are not the same. If you look only at the total bill, you cannot see which item is growing.
Six things to measure
- Input and output tokens, separately. Because output and input are priced differently, total tokens alone mislead.
- Which model. Record which model served each task; giving easy work to a smaller model is one of the biggest levers.
- Cache hit rate. If a long repeated system prompt is served from cache, its cost drops as a separate item; track the hit rate.
- Retries and errors. Failed and repeated requests also spend tokens. A per-request average hides that.
- Tool calls and extra steps. If the model loops several turns per answer, one user question splits into several requests.
- Business outcome. Questions answered, summaries produced, tickets closed. Without dividing cost by outcome you cannot tell whether it is cheap or dear.
The right unit: the job, not the request
'Average cost per request' answers one question: how much am I spending? The real question is 'what does it cost to finish one job?' The cost of one report summary, of one draft reply to a customer e-mail, of one document classification. Those units make sense to the business owner and let you decide whether to keep a feature.
When you log
- Record counts, not amounts. Store token counts, model name and time; compute the amount at report time with that day's price table. When prices change, old records stay intact.
- Log numbers, not text. Prompts and replies may contain personal data; measurement does not need them.
- Set a ceiling. Raise an alert as you approach the daily or monthly budget; an uncontrolled loop can burn many tokens overnight.
Where Floemdata stands
Floemdata's AI assistant is on the roadmap in the product documents; the code has no section for it yet. We therefore did not find any mechanism inside Floemdata that measures AI usage cost. This article is a general measurement plan; it does not describe a feature of the product today.
To begin, open a table with these columns: date, model, input tokens, output tokens, task type. After a week you will have the data to ask your first question. Questions are welcome at info@floemdata.com or via Floemdata.