What a call costs, and how a loop becomes a bill

The two pricing worlds

A subscription is a flat monthly fee per person, typically USD 20 to USD 30, for a chat window with limits you will rarely hit. Predictable, and covered in module one.

Usage pricing is what you meet the moment you automate. You pay per token, separately for what goes in and what comes out, and there is no ceiling until you build one.

A token is roughly three-quarters of a word in English, and rather less efficient in most other scripts, a point worth knowing if you work in Hindi, Tamil or Arabic, where the same sentence can cost noticeably more tokens than its English translation.

Work an actual example

Say you classify and draft a reply to each customer enquiry. Per enquiry:

• Facts sheet and instructions sent in: about 900 tokens.
• The customer’s message: about 200 tokens.
• The drafted reply out: about 300 tokens.

At the prices of a small, fast model, take USD 0.15 per million tokens in and USD 0.60 per million out as a realistic order of magnitude: that is about 0.00016 USD in and 0.00018 USD out. Roughly USD 0.0003 per enquiry.

Three hundred enquiries a month is about nine US cents. Not a typo. For ordinary small-business volumes, the model is not the expense; your time and the subscription are.

Use a large frontier model instead, at perhaps twenty times the price, and it becomes about USD 2 a month. Still not the expense.

So where does the money actually go

Three places, and none of them is the per-enquiry cost.

Resending context. You pay for the input every single time. That 900-token facts sheet is charged on every one of the three hundred calls. Double the facts sheet and you double the input bill without doing any more work. This is why the length of your standing instructions is a cost decision, not just a clarity one.

Long conversations. Module two explained that the whole thread is re-read on every turn. The cost consequence is that a twenty-turn conversation costs far more than twenty single questions: turn twenty pays for turns one to nineteen again. Prompt caching, offered by most providers, cuts the repeated portion substantially and is worth switching on, but the shape of the problem stays.

Loops. This is the one that produces the alarming invoice.

How a loop becomes a bill

An automation that calls the model, and whose output triggers something that calls the model again, can run without limit. A retry that does not check whether the first attempt succeeded, a rule that files a document and a rule that processes filed documents, an agent that decides to try once more: each is an ordinary bug, and each can run thousands of times overnight.

The per-call cost is what makes this dangerous rather than merely embarrassing. At a hundredth of a penny per call, nobody notices ten thousand calls until the statement arrives.

Three defences, and set all three before the first automated call:

1. A hard spending cap on the account. Every major provider has one. Set it to a number that would annoy you rather than hurt you: for most small businesses USD 20 or USD 50 a month is generous.
2. A separate key for each automation. When one misbehaves you can revoke it without stopping everything else.
3. A counter. Anything automated should refuse to run more than N times an hour, where N is three times your realistic maximum.

Rate limits, which are not costs

Separately from money, providers cap requests per minute and tokens per minute. Hit the cap and you get an error, not a bill.

This matters for batch work: firing two hundred product descriptions at once will fail partway through. The fix is a pause between calls and a retry with an increasing wait, but read the next lesson before writing any retry, because a naive one causes the duplicate problem.

The honest comparison

Free tiers exist on most providers and are real. They are also rate-limited, sometimes routed to smaller models, and frequently carry different terms about whether your inputs can be used for training. If the work is customer data, read those terms before the price.

The free path is running a small open model yourself, covered in module one. The per-call cost is then zero and the cost is your hardware and your attention. For a business doing three hundred enquiries a month, paying nine cents to somebody else is usually the better trade; the free path earns its place when confidentiality, not money, is the binding constraint.