How are you pricing your AI agents, and do you know what a run actually costs?

I keep running into the same thing with people building AI agents and AI products: everyone can quote the headline price of a model, but almost nobody can tell you what one full agent run costs them.

The reason is that an agent isn't one call. It's a planning step, a few tool calls, some retries, maybe a summary, sometimes across two or three different models. The cost stacks across steps and usually concentrates in one step you didn't expect. Then a provider changes its rates and the math you did last month is quietly wrong.

Which feeds the harder question: how do you even charge for it? Per seat doesn't fit. Usage, outcome, and credit-based pricing all have tradeoffs, and which one works depends on numbers most builders never actually work out.

So I'm curious how this community handles it:

• Do you know your real cost per run, or do you estimate?

• Usage, outcome, credits, or hybrid, and why?

• Has a model price change ever wrecked your margin without you noticing?

For context, I'm building PriceKit to answer the cost-per-run part, so I'm not neutral, but I'm genuinely trying to learn how people approach the pricing side. What's working for you?