The first question in any meeting is the same one: “so what does it cost per month?” It is the right question, and it rarely has a short answer, because an agent is not billed the way the software you already own is billed.
An ERP is billed per user. An agent is billed per unit of work. Double your headcount and the ERP costs twice as much while the agent costs the same. Double your quote volume and the ERP costs the same while the agent costs twice as much.
That changes how you budget, and it also changes which design decisions cost you money. Here is the full breakdown, with the arithmetic done.
The billing unit is the token
A token is a fragment of a word. Providers charge per million tokens, and they charge separately for what goes in and what comes out.
Three things worth knowing up front, because these are what break budgets estimated by feel:
The same text is not the same number of tokens in two different models. Tokenizers change between generations: Anthropic’s own documentation warns that recent models produce roughly 30% more tokens for the same text. A budget you calculated with last year’s model does not carry over by itself.
Every call pays for everything you send it, again. The model does not remember the previous call. If your agent pastes the full catalog into every step and takes twenty steps, you paid for the catalog twenty times. This is the number one cause of invoices that look nothing like the estimate.
The price between models varies a hundredfold. On Anthropic’s public list today, the cheapest model takes input at $0.10 per million tokens and the most expensive at $10. Picking the model is not a technical decision, it is a cost decision.
The answer costs five times the question
This is the asymmetry almost nobody accounts for, and the easiest one to use in your favor.
Across the whole price list, output costs five times input. A model at $2 per million input tokens charges $10 per million output. One at $4 charges $20. One at $0.10 charges $0.50.
The practical consequence: an agent that explains costs far more than an agent that answers. Ask it to fill six fields and you pay for six fields. Ask it to fill six fields and also narrate its reasoning in three paragraphs, and you pay for the six fields plus the three paragraphs, at five times the price of the input you sent to produce them.
Asking for a short, fixed output format is not a programming detail. It is, without exaggeration, the largest cost lever available after choosing the model.
One real task, costed end to end
Take the task we see most often: receive a request for quote by email and leave the quote prepared for a person to review.
What goes in: the email, the parts list, whatever was found in your quoting history, and the current material cost. Call it 15,000 tokens, which is a long email with a table and context.
What comes out: the drafted quote. Call it 1,500 tokens.
With a mid-range model at $2 input and $10 output:
- Input: 15,000 × 2 ÷ 1,000,000 = $0.03
- Output: 1,500 × 10 ÷ 1,000,000 = $0.015
- Total: $0.045 per quote. Four and a half cents.
Twenty quotes a day, twenty-two working days: 440 a month. Under $20 a month.
And since a real agent does not take a single step — it reads the email, searches the history, drafts the reply — multiply that by three calls with accumulating context and you land at roughly $48 a month.
Now put the other number next to it. If each quote is assembled by hand in 25 minutes today, those 440 quotes are 183 hours a month. More than a full salary spent transcribing, searching and copying.
That is the comparison that matters, and it is why the model invoice is almost never the problem. The model is the cheap part.
The three discounts almost nobody uses
If your estimate comes out much higher than this, you are probably leaving money on the table in one of these three places.
1. Prompt caching. If every call starts with the same instructions and the same catalog, that part can be cached. Reading from cache costs between a tenth and a twentieth of the normal input price. Writing it costs a little more than normal input, so the trick is to put the fixed part first and never change it. On an agent that takes several steps over the same file, this alone cuts the invoice in half.
2. Batch mode. If you do not need the answer this second, batch processing costs half, on both input and output, in exchange for being processed within a window of hours. Everything that runs overnight — reviewing tomorrow’s schedule, chasing documents for tomorrow’s shipments — belongs there.
3. The small model for the dumb part. Classifying an email, extracting a parts list or deciding whether a message is a request does not need the expensive model. Save the expensive one for the step that drafts something a customer will read. Between the economy model and the premium one there is a hundredfold difference in input price, and on a classification step you will not notice the quality difference.
All three stack. An agent designed with all three and one designed with none can differ by an order of magnitude doing exactly the same work.
What the vendor does not invoice, and it is most of the cost
This is where budgets actually go, and none of it shows up on the model invoice.
Integration with your systems. It is the main cost of any useful agent, and it has nothing to do with artificial intelligence: it is reading your quoting history, looking up your material cost, writing into your system. If the data lives in three people’s Outlook folders, the work is getting it out of there, and that work bills by the hour like any other development.
The owner of the exceptions. The agent will get stuck, and someone has to receive what it could not resolve. That is hours of a person every week, and hours of someone who knows the subject, not someone cheap. If you do not budget them, you will pay them in the form of a project that quietly dies after two months.
Maintaining the instructions. The prompt is not a config file you write once. The process changes, a customer’s format changes, a new model ships, and it all has to be retested.
We say this from our own experience. On our diagnostic agent we tightened the output token budget to save money, and two of the fifteen areas started returning errors without a single test catching it, because the tests simulated the model’s response instead of asking for it. Saving in the wrong place cost us days of a service that was broken silently.
And a technical detail worth money: the output token limit is a ceiling, not a charge. Lowering it saves you nothing if the answer already fit; all it does is truncate the answer when it does not. To actually save, you have to ask for shorter answers, not merely allow them.
Testing costs money. Every time you test the agent end to end, it bills. Our full test run costs about thirty cents, which is why it sits outside the automated checks and requires an explicit flag to run. If you plan to test seriously, that goes in the budget.
Set the cap before you turn anything on
An agent is software that calls a paid API in a loop. A bug in that loop does not take down the server. It bills.
Before exposing anything publicly, two caps:
- One per user per day, so that nobody — a person or a bot — can spend your month in an afternoon.
- One on attempts per task, so a failing step does not retry forever.
Our public agent has both, and the first exists precisely because the page is open and anyone can use it.
The four numbers you need before asking for a quote
Walk into a vendor meeting without these four and every price will sound reasonable and none will be comparable.
- How many times a day the task happens. It multiplies everything else.
- How much text goes in and how much comes out, in a typical case. An email and a table, or a forty-page file. The difference is two orders of magnitude.
- How long a person takes today, and the loaded cost of that hour. It is your baseline and your only measure of success.
- What happens when it gets it wrong. If every output needs someone to review it, that review is part of the cost. Risk is paid in review hours, not in model dollars.
With those four numbers you can do the arithmetic above yourself, in a spreadsheet, before anyone quotes you.
What you should not be paying for
Do not pay per seat. An agent does not sit down. If you are charged per user per month for something that runs on its own without anyone opening a screen, ask what exactly you are buying.
Do not pay for the label. “AI-powered” on top of an email rule that already existed is not an agent. The test is simple: ask what happens when the buyer sends the part number in an attached PDF instead of in the subject line. If the answer is that it breaks, it is ordinary automation and it should cost what ordinary automation costs.
Do not pay for tokens you do not use. If a vendor sells you a monthly capacity bundle, ask for the actual consumption detail. Usage-based billing is the natural shape of this, and a fixed bundle is almost always sized in favor of whoever sells it.
Where to start
Pick one task, measure the four numbers, and do the arithmetic. If the model costs tens of dollars and the task consumes hundreds of hours, you have a case. If it comes out the other way, you just saved yourself a project.
And if it is still not clear what counts as an agent and what does not, start here: the difference between an agent and a chatbot is exactly what decides whether the thing being quoted to you does work or merely replies.
Before automating, it is also worth checking whether this is the right fight. An agent inherits the quality of your data, and if inventory or costing are shaky, some things come first.
If you want those four numbers gathered properly and in writing, that is what Chira is for. Five minutes, it asks about volume and about where the data lives, and you leave with a document you can take into the meeting. You do not end up on a sales call.
References
The prices and discounts in this article come from Anthropic’s public list as of the publication date. Model prices change, so verify the number before budgeting with it. What does not change as fast is the structure: billed per token, output more expensive than input, caching and batching discounted.
- Claude API pricing, Anthropic — the per-model table, cache pricing, the 50% batch discount and the note about the new tokenizer.
- Batch processing, Anthropic — how the asynchronous window behind the discount works.
- Prompt caching, Anthropic — which part of the prompt can be cached, and what writing and reading it costs.
- Amazon Bedrock pricing — useful for comparison if your cloud is already on AWS.