AI & LLM cost
Model tiering
Routing each request to the cheapest model capable of handling it well, instead of sending every request to the most capable (and most expensive) model available.
Last updated
Definition
Not every request needs a frontier model: classifying a support ticket's category, extracting a field from a form, or checking a simple condition can often be handled by a smaller, cheaper model just as reliably as by the most capable one, while a hard reasoning task can't. Model tiering means explicitly routing requests by difficulty, using cheap models for the bulk of easy cases and reserving expensive ones for the fraction that needs them.
Done well, this is usually the single biggest lever on blended AI cost per request, larger than any individual prompt optimization. It changes the rate paid on the majority of volume, not just the efficiency of any one call.
Where it shows up
Model tiering is implemented in application code that routes a request to one of several models by name, each with its own published per-token rate on the provider's pricing page, rather than sending every request to the same model regardless of what the task actually needs from it.
What makes it expensive
The trap here is sending every request to the most capable model available regardless of task difficulty, paying its top rate on volume a cheaper model could handle just as reliably. A blended cost per request that holds flat as volume grows means nothing is being routed at all. Broken 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.