What does a single run cost?

Codified steps run as plain code with no model call, so they cost fractions of a cent. An agent step costs whatever the model usage costs. A human approval costs nothing but the time of the person answering it. Every run records what it spent, so the bill at the end of the month is a sum of numbers you have already seen.

That is why the price per run falls as the workflow is learned. The same job that needed an agent for twelve steps ends up needing one, and the eleven others became code that runs in milliseconds.

Learning runs are the expensive ones. They are agent heavy on purpose, and there are only a handful of them before the workflow is operationalized.

How execution is billed is your choice: metered usage, your own model key with no markup on tokens, or a price per successful outcome agreed per workflow before it goes live. The platform subscription is separate and covers everything that is not execution.