LLM cost calculator.

What actually drives the bill

Teams tend to agonise over per-token pricing between models and then get surprised by the invoice anyway. The model choice is usually the smallest of the four levers below.

Conversation length, not message count

Most chat implementations re-send the entire thread on every turn, so cost grows with the square of conversation length rather than linearly. Truncating or summarising history is usually a larger saving than switching models, and it's entirely within your control.

Retrieved context is input you pay for

A RAG system that stuffs ten documents into the prompt pays for ten documents on every single query. Reranking to the best three usually improves answer quality and cuts input cost at the same time — one of the rare changes that helps on both axes.

Caching changes the maths on repeated context

If a long system prompt or document set is re-sent on every call, prompt caching can reduce the cost of that repeated portion substantially. It matters most in exactly the high-volume workloads where the bill is worth worrying about.

Not every step needs the strongest model

Classification, routing, and extraction often run well on a smaller, cheaper model, with the frontier model reserved for the reasoning step that genuinely needs it. Mixed-model pipelines are how production systems keep cost sane without giving up quality where it counts.

These are estimates built from published list pricing, so treat the output as a planning figure rather than a quote — and remember that inference is only part of running cost. Human review time and eval maintenance belong in the same budget.