The True Cost of One LLM Call
Tokens, latency, and dollars; the real math behind a single call.
Every call to a model has a price, and it is made of tokens and time. Once you can do the math on a single call, the cost of a whole system stops being a mystery.
What you pay for
A call is billed by tokens, in two directions: the tokens you send in (your prompt, the conversation, any documents) and the tokens the model writes back. Output usually costs more per token than input. So a long prompt with a short answer and a short prompt with a long answer can cost very different amounts.
The third cost: time
Tokens cost money; they also cost latency. The model writes its answer roughly one token at a time, so a longer answer takes longer to arrive. For anything interactive, that wait is part of the cost too.
Doing the math
Take the input tokens times the input price, plus the output tokens times the output price. That is one call. Multiply by how many calls your feature makes (an agent that loops ten times makes ten calls) and you have the real cost of a single user action. Most cost surprises come from loops and from stuffing too much into the prompt.
Why it matters
Once you can price one call, you can reason about a whole system: what a feature costs per use, where to trim the prompt, when to cache, whether an agent's extra loops are worth it. Cost stops being a scary monthly bill and becomes something you design around.