AI & LLM cost
Model tiering
Routing each request to the cheapest model that can handle it well, not sending every request to the most capable and most expensive model.
Last updated
Definition
Not every request needs a frontier model. A smaller, cheaper model can often classify a support ticket, extract a field from a form, or check a simple condition just as reliably. A hard reasoning task cannot be handled that way. Model tiering means you route requests by difficulty. Cheap models take the bulk of the easy cases. Expensive models take the fraction that needs them.
Done well, this is often the biggest lever on blended AI cost per request. It can be larger than any single prompt change. It changes the rate you pay on most of the volume, not just the efficiency of one call.
Where it shows up
Model tiering lives in application code. The code routes a request to one of several models by name. Each model has its own published per-token rate on the provider's pricing page. Without tiering, every request goes to the same model, whatever the task needs from it.
What makes it expensive
Some teams send every request to the most capable model, whatever the task difficulty. They pay its top rate on volume that a cheaper model could handle just as reliably. If the blended cost per request holds flat as volume grows, nothing is being routed. Break cost out by task type. The easy majority becomes obvious.
Related
Token spend moves every day.
Model prices and rate limits change between invoices. CostMon tracks your AI spend daily, so a change shows up before the bill does.