What an AI Agent Actually Costs to Run: The Math Nobody Shows You

Every agent proposal has the same silence in it. The build cost is right there on the quote. The running cost is somewhere between a shrug and “usage-based,” which reads like a warning label. So founders quietly assume the meter will bite them, the way cloud bills once did, and shelve workflows that would have paid for themselves in a quarter.

Here is the actual math. It is three numbers multiplied together, you can do it on a napkin, and for most business workflows it lands somewhere surprising: the running cost is the smallest number in the room.

The whole formula on one line

An agent’s marginal cost is: tokens per task x price per token x tasks. That is the entire meter.

A token is the unit AI is metered in, roughly three-quarters of a word, priced per million. Mid-2026 blended rates run from under a dollar per million tokens on budget models to a few dollars mid-tier and tens of dollars at the frontier. The numbers move, the structure does not.

So take a real task: an agent that reads an inbound support email, pulls the relevant policy and the customer’s history, and drafts a grounded reply. Everything it reads and writes might total 20,000 tokens with a retry included. On a mid-tier model at an illustrative $4 per million, that is 8 cents. The person who used to do that task took 20 minutes, and at a loaded $45 an hour, that same task cost $15. Not 20 percent less. A different unit of measurement.

Run your own numbers:

Play with the extremes and notice how hard it is to make the agent expensive. Even at frontier pricing and heavy token counts, the per-task cost rarely climbs out of pocket-change territory while the human line sits at dollars. This is why the “model was never the bottleneck” shift happened: intelligence became a metered utility, and the meter is cheap.

Where the tokens actually go

The bill is not evenly spread, and knowing its anatomy is how you keep it small.

Reading dominates. The context an agent takes in, documents, history, instructions, is usually several times larger than the answer it writes. Which means cost control and answer quality are the same discipline: engineering the context so the model sees the three passages that matter instead of forty documents that might. Sloppy retrieval is the most common reason a bill runs hot, and it makes answers worse at the same time.

Multiplication hides in architecture. A single agent that loops five times, or a multi-agent setup where several agents each read the full context, multiplies tokens in ways no line item shows. This is not a reason to avoid those designs. It is a reason to know your tokens-per-task number, not just your price-per-token number.

Model tier is a dial, not a destiny. Most routine steps run fine on mid-tier models; save the frontier rates for the judgment-heavy step that needs them. Mixed-tier designs routinely cut the bill by half or more with no visible quality change, and your evals are what prove the “no visible quality change” part instead of hoping it.

You do not manage an agent’s cost by watching the bill. You manage it by designing what the agent reads.

The honest caveats, before you over-celebrate

Tokens are the marginal cost, the cost of the next task, and that is the number that makes scaling decisions easy. It is not the whole economics. The build is real money up front. Hosting, monitoring, and evals are real money monthly. And saved minutes only become value when the time actually goes somewhere, which is the realization problem the ROI calculator models explicitly, and the reason that post’s default scenario comes out negative in year one. Cheap to run has never meant automatically worth building.

The right way to hold both numbers: the ROI math decides whether to build. The token math tells you what happens after it works, and its answer is the strategic one: scaling from 40 tasks a day to 400 raises the agent’s bill by a few hundred dollars a month, while doing the same by hand would take most of a team. Once a workflow is automated and trusted, volume is nearly free. That asymmetry, not the pennies, is the real headline.

The other meter, running the whole time

The same tasks, done by hand

$118,800

$9,900 a month, month after month

12 months · $118,800 6 months · $59,400 24 months · $237,600

What to do with this

Pick the one repetitive, document-heavy task your team complains about most, and put its three numbers into the calculator above: how many times a day, how many minutes by hand, and a guess at the tokens. If the ratio comes out where most workflows do, somewhere between 50x and 500x, the running cost was never the obstacle. The obstacle is the build, the trust, and the rollout, which are all solvable, scoped problems with a known first step.

---

Want the real number for your workflow instead of an illustration? That is a one-conversation estimate. Book a free consult and we will size the tokens, the tier, and the build for one agent worth having, with the running cost in writing.