The real cost of per-token pricing

A client came to us paying roughly $4,200/month on a per-token API for a support-ticket triage workflow. At their current growth rate, that number was on track to hit $11,000/month within a year — for a workflow that runs the same handful of prompts against the same handful of ticket types, over and over.

Per-token pricing makes sense when your usage is unpredictable or exploratory. It makes much less sense once a workflow is stable and high-volume — at that point you're paying a variable-rate premium for something that behaves like fixed infrastructure.

We moved them to a self-hosted deployment with fixed monthly hosting. Their cost curve flattened immediately, and it stayed flat as volume grew. The lesson isn't "self-host everything" — it's "know which of your workflows have graduated from experiment to infrastructure, and price them accordingly."