The full answer
LLM usage is billed by tokens, the units of text sent to and returned from a model, so cost depends on how much context each request carries and how often requests happen. Design decisions made early, such as how many documents to retrieve or how long a system prompt is, have a large effect on the monthly bill.
Model routing is the biggest lever. Classification, routing and simple extraction can run on small, inexpensive models, while complex reasoning goes to a frontier model only when needed. Creuto builds this routing into the application and verifies with the evaluation set that quality holds.
Caching avoids paying twice: identical or near-identical requests return stored results, retrieval results are reused within a session and provider-side prompt caching is used where available for long, stable context.
Operational guardrails include rate limits, per-tenant budgets, alerts when spend deviates and dashboards that show cost per feature, per customer and per model, so you can see which features earn their keep.
For very high volumes, a self-hosted open-weight model can be cheaper than API pricing, and Creuto will model that break-even point for you.