What an AI agent actually costs to run, priced line by line

In short: most agent products are sold per seat or per month, so the price is disconnected from what the agent does. Priced honestly, a run is: the model calls it makes, plus a compute charge that is a fraction of a cent, plus a per-call fee on the handful of tools that cost real money. Creating and editing agents is free, most connector calls are free, and a per-run cap stops the pathological case. The number that matters is not the monthly fee — it is the cost of one run, multiplied by how often it fires.

Why per-seat pricing hides the thing you need to know

A seat licence answers "how many people may open the dashboard". Nobody is asking that. The question is what happens to the bill when an agent that runs hourly starts taking twice as many steps, or when a customer-facing task suddenly gets used.

Per-seat pricing cannot answer it, so vendors add usage caps underneath, and the caps become the real price — discovered when you hit one. Usage pricing is less comfortable to put on a pricing page and it is the only shape that tells you what an automation costs before you switch it on.

The parts of a run

Two things in that table matter more than the numbers.

Editing is free. If iterating on an agent costs money, you stop iterating, and you keep a worse agent because improving it has a price. The compile is what costs, because it calls a model.

Most connector calls are free. Reading your Shopify orders or writing a row to your own spreadsheet has no per-call fee, because it costs us nothing per call. The exceptions are the ones where somebody else charges us — X charges per API call, so that is passed through at twice their rate, stated plainly rather than buried.

The number to actually compute

Take the agent you are considering and work out one run. Roughly:

> (planning calls × model rate) + (a fraction of a cent) + (priced tool calls × their rate)

Then multiply by frequency. An hourly agent runs 720 times a month; a daily one runs 30. That factor of 24 dominates every other decision you will make, and it is entirely under your control — most "hourly" automations are answering a question that changes daily.

A trivial reporting agent making a couple of model calls costs a small fraction of a cent in compute and the price of those calls. An agent that generates an image costs what the image costs, which is the dominant term the moment one appears. Model rates are on the rates page, per model, because the model you choose changes the answer more than anything else here.

Where the money actually goes wrong

Not in the per-call rates. In three places:

Frequency nobody revisited. An agent set to hourly during testing, left there. This is the most common and the easiest to fix.

Retries on a broken dependency. A connector that fails and a run that keeps trying. The per-run cap is the backstop; a run that reaches $2.00 stops rather than continuing to spend.

A priced tool inside a loop. Web search or image generation called once per item over a long list. Worth reading the step log of a real run before scheduling it, precisely because that is where you see it.

What should happen when you run out

Failing loudly, not partially. Compiles are refused, runs fail with no_credits, and a customer who invokes an agent from chat is told it is unavailable rather than getting half an answer. Runs that fail because of a platform fault are credited back.

That last one is worth asking any vendor about. An agent that failed because *their* infrastructure broke and still billed you is a pricing model with a bug in it, and you will not notice it on any single invoice.

The honest summary

Agents are cheap per run and expensive per schedule. The cost of one run is usually a fraction of a cent to a few cents. The cost of a decision to run something hourly instead of daily is 24× that, forever.

So the useful discipline is not choosing a cheaper model. It is asking, for each agent, how stale an answer you can tolerate — and then setting the schedule to that, rather than to the fastest thing the interface offers.

FAQ

How much does it cost to run an AI agent?

Priced per run rather than per seat, a run is the model calls it makes plus a compute charge that is a fraction of a cent even for a heavy run, plus a per-call fee on the few tools that cost real money. A simple reporting agent is a fraction of a cent to a few cents per run; an agent that generates an image costs what the image costs. The model you pick changes the answer more than anything else — per-model rates are on the rates page.

Is per-seat or usage-based pricing better for AI agents?

Per-seat pricing answers how many people may open the dashboard, which is rarely the question you have about an automation. It also cannot tell you what a run costs before you schedule it, so vendors add usage caps underneath and those caps become the real price — usually discovered when you hit one. Usage pricing is less comfortable to publish and is the only shape that prices the thing you are about to trigger.

What makes AI agent costs unexpectedly high?

Three things, none of them the per-call rate. A schedule left on hourly after testing, which is 720 runs a month against 30 for a daily one. Retries against a broken dependency. And a priced tool such as web search or image generation called once per item inside a loop. Reading the step log of one real run before scheduling it catches all three.

Do I pay to create or edit an AI agent?

Adding one from a library, duplicating, editing and changing a schedule are free. Compiling an agent from a written brief costs the model's own price for that compile and nothing on top. Charging for edits would mean people stop iterating and keep worse agents, which is the wrong incentive to build in.

What happens if my balance runs out mid-task?

It should fail cleanly rather than partially. Compiles are refused, runs fail with no_credits, and a customer invoking an agent from chat is told it is unavailable rather than receiving half an answer. Runs that fail because of a platform fault are credited back — worth asking any vendor about, because a model that bills you for their own outage has a bug you will not spot on a single invoice.

*Per-run costs appear in the run log, and each agent's month-to-date spend is on its row in the Agents tab. The Agent Tasks guide has the full breakdown and the limits.*