Price the agent before you build it
In short: A tool that prices one representative agent run and splits it across the six lines from the previous article. It shows two things a table cannot: how small a share of the bill tokens usually are, and which of the two levers — a cheaper model or a simpler architecture — saves more on your own numbers. Rates are list prices, the maths runs in your browser, and nothing leaves the page.
The previous article on what an AI agent actually costs ends with advice that is hard to follow: price one representative task end to end, and only then multiply by volume. The advice is right and nobody takes it, because it means holding six lines in your head at once and multiplying one of them by another.
So here it is as a tool.
Runs in your browser. Nothing is sent anywhere.
1 · One representative run
2 · Architecture and volume
3 · What comes back
4 · People who process the output
5 · The harness over the output
Tokens as a share of the total —
Which lever saves more
On your numbers the model change saves more than the architecture change — that is the rarer case.On your numbers the architecture change saves more than the model change. That is what the article above is about.On your numbers the two levers come out about even.
Rates are list prices as of the day you read this; the calculator does not fetch them, so check them before you quote them anywhere.
Where the numbers come from
The calculator is exactly as good as the worst number you put into it. Three of them can be measured, the rest can only be estimated honestly — and knowing which is which matters more than the arithmetic.
Input and output tokens per run are measured, not estimated. Run one real task, not a shortened demo, and read the actual usage off the API response. An estimate along the lines of "call it twenty thousand tokens" is usually off by an order of magnitude for an agent, because every step carries the whole accumulated context with it.
The architecture multiplier is a design decision, not a measurement. That is why it sits in its own field: tokens per run change when the task changes, the multiplier changes when the design changes. The options are Anthropic's published measurements — roughly four times chat for a single agent, roughly fifteen times for a multi-agent system.
The share of input that can be cached is the stable part of the prompt: the system prompt, a reference document, the tool list. Mind the trap — caching is a prefix match, so a single timestamp in the system prompt drops the whole discount and nothing warns you. Before you type 80% in here, confirm on a real run that the cache is actually being hit.
Minutes of review is the number people revise downward. Do not ask how long the output takes to read. Ask how long it takes to act on — including the cases where it is wrong and nobody finds out for twenty minutes.
The cost of the harness is building the thing that checks the output: a judge, tests, an eval set, pairing against known-good answers. Amortise it over as many months as you expect the system to run unchanged. Anyone typing sixty in that field is assuming that in five years neither the model, nor the process, nor the price list will move. All three will.
What it deliberately leaves out
Currency conversion. Model rates are billed in dollars and the invoice arrives in dollars. An exchange rate is a number that moves daily, and hardcoding one into a page means being wrong about it within a month.
Differences in model capability. The calculator can compare what Haiku costs against Opus. It cannot tell you whether Haiku can do the work. That is a question for an eval, not a table, and a cheaper model that fails to finish the task is infinitely expensive.
Live rates from the network. The calculator does not fetch anything — it uses list prices as of the day you read this, with one exception it does handle: the introductory rate for Sonnet 5, which holds until 31 August 2026. After that day it switches itself to the list rate, because the browser knows the date. For everything else the rule stands: check the rate before you quote it anywhere.
What you see when you fill it in honestly
Two things, and both are why the tool exists.
Tokens are usually the minority of the bill. On the default inputs — two thousand runs a month, six minutes of review per output — tokens come out at roughly a tenth of the total. The rest is people and harness. Optimising only that tenth means tuning the smallest number on the screen and leaving the other ninety per cent alone.
Architecture moves the invoice more than the model does. The calculator computes both levers side by side from the same inputs: one model class cheaper against one architecture tier simpler. Architecture wins on most workloads, and that is not an accident — the spread between model classes is a small multiple, the spread between a chat call and a multi-agent system is fifteen.
Try your own numbers against that claim, though. There are workloads where the model wins: short runs, high volume, no review. The calculator will notice and say so — it has an arithmetic built in, not a conclusion.
One last note
The calculator runs in the browser and sends nothing. Volumes, hourly rates and cost estimates do not leave the page — not only because that would be impolite, but because this site's security policy does not permit external scripts at all. If it produces a number you would rather not show anyone, you do not have to: nobody else can see it.
And if it produces a number that only works in the cheapest model class with the tightest settings, you have the answer the exercise is for.
Frequently asked questions
How do I estimate what an AI agent will cost?
Price one representative run end to end, then multiply that by volume. The calculator on this page adds the four things estimates usually lose: the architecture multiplier, what caching and batching give back, the hours of the people who process the output, and the amortised cost of the harness that checks it. A per-token estimate without those four lines is usually less than half the real invoice.
How many tokens does one agent run use?
That cannot be estimated, only measured — run one real task and read the actual usage. The calculator asks for tokens per run and the architecture multiplier separately because the first is a measurement and the second is a design decision, and they change for different reasons.
Is switching to a cheaper model worth it?
Almost always less than people expect. The spread between model classes is a small multiple; the spread between a chat call and a multi-agent system is fifteen. The calculator computes both levers side by side from the same inputs, so you can see which one is bigger for your workload — and for most workloads it is not the model.
Does the calculator send my numbers anywhere?
No. Everything is computed in the browser, there is no external script on the page, and neither volumes nor hourly rates leave it. The site's security policy does not permit external scripts at all.
About the author
Filip has spent his career between media and technology — filming for ŠKODA, Pilsner Urquell and Range Rover, then co-founding a startup in San Francisco, working with a YC-backed company and serving as CIO at Renato. He built ZEUS Legal AI and now runs the systems Salamon Capital operates on.
New analysis, when there is one
One message when a new analysis or article goes out. Nothing else — no sequence, no digest, no third party. The address is stored so there is somewhere to send it, and one click deletes it for good.