What an AI feature really costs, and what drives the price

When people ask what an AI feature costs, they usually mean the price of building it. That is only the first bill. The second arrives every month for as long as the feature is in use. Whether you’re approving the budget or wiring up the feature yourself, you need both numbers, and you need to know what moves them.

Two price tags: building and running

Building cost is the work to design, write, test and launch the feature. You pay it mostly once, plus improvements later.

Running cost is what it takes to keep the feature working, every day:

• Model usage. AI providers charge per request, based on how much text goes in and comes out. They count it in tokens, small chunks of text a few characters long. Think of a taxi meter: the longer the trip, the higher the fare.
• Hosting. The servers and databases the feature runs on.
• Monitoring. Tools that watch for errors, slow responses and unusual costs.
• People reviewing. Someone checking a sample of answers, handling the cases the AI can’t and improving it over time. This is often the largest running cost, and the one most often left out of the plan.

What makes it cheap or expensive

Four things move the price more than anything else.

• Volume. One cent is nothing once and real money a million times. Estimate how many requests you expect a day, then check the estimate after launch.
• Model size. Bigger models are usually better at hard tasks and cost more per request, sometimes many times more. Smaller models are cheaper and faster, and good enough for simple jobs like sorting messages or short summaries.
• Context size. Context is everything you send the model with each request: instructions, documents, conversation history. Sending a whole manual when one page would do means paying for the whole manual, every time.
• Quality bar. How good must the answer be? A rough draft that a person will edit can come from a cheap model. An answer that goes straight to a customer, or to a child, needs more care, and usually costs more.

Estimate the cost of one useful result

A monthly bill is hard to judge on its own. A better number is the cost per useful result: per answered question, per processed document, per finished lesson. Count everything, including the requests that fail or need a retry, and the minutes a person spends checking.

A real example. On SuperEscuela, a free learning platform for children, AI writes the lessons. Each lesson costs about 2 to 4 cents in AI with the higher-quality model; the lessons for a whole school come to roughly $60 to $130. We tried the smallest model first, because it was cheaper. A quality review showed its lessons weren’t good enough for children, so we replaced it. The cheapest model per request wasn’t the cheapest per useful lesson, because a lesson that has to be rewritten isn’t useful yet.

That is the pattern to copy: get the quality right first, measure it, then look for savings that keep the quality where it is.

Set budget alerts before you launch

AI costs rarely creep up slowly. They jump: a bug that repeats a request in a loop, a page that suddenly gets traffic, a stranger using your form as a free AI service. Protect yourself before it happens:

• Set a monthly spending cap with each AI provider, and an alert well before you reach it.
• Limit how many requests one user can make per minute or per day.
• Track the cost per useful result every week, not only the total.
• Name the person who receives the alerts and is allowed to act on them.

Questions to ask

Ask your team, your vendor or your AI tool: What is the expected cost per useful result? Which model are we using, and why that one? What are we sending with each request, and does all of it need to be there? What happens to the bill if usage grows ten times? Clear answers mean the cost is understood. A shrug means it isn’t yet.