Token price is not total cost: the real AI operations budget

Model providers commonly price input and output volume. The rate is useful, but it does not reveal how many calls are required for one accepted task, how context grows or how often failed attempts repeat.

OCR, search, databases, infrastructure, observability, backup, security tooling and third-party licences are separate. A provider or model change can also trigger re-testing work.

Use a completed task that passes acceptance criteria as the denominator—not tokens. If a document takes three attempts and one in ten cases needs human intervention, all of that belongs in the transaction cost.

Route simple work to smaller models, clean long context, cache reusable results and cap the number of steps. When a budget threshold is reached, alert or stop rather than spending silently.

Refresh price comparisons at proposal time. Today’s provider rate is not a long-term fixed-cost guarantee.