Inference cost
Also called inference · usage cost · API cost
What a provider charges you to actually run a model, billed per unit of work done, not per month.
In more detail
Inference is the act of running a trained model on your input to get an output. Providers charge for it by usage: per token for text, per second or per character for speech, per image for stills. The defining property is that it scales with what you produce, so an idle month costs nothing and a heavy month costs proportionally more.
Example
The same script costs the same whether it is your first of the month or your fortieth, so the monthly figure is just that unit price times volume. This is arithmetic you can do before committing, which is not true of a subscription tier that caps you at an arbitrary number of generations.
Why it matters
It is the number that decides whether owning a stack is actually cheaper than renting one, and the honest answer depends entirely on volume. Below some publishing rate a subscription wins on price and convenience; above it, usage billing wins and keeps winning. Anyone who tells you one is always cheaper has not done the multiplication.