How much does it cost to use AI in your business: tokens, APIs and the hidden costs
Using AI in a business has two very different costs: building it, paid once, and using it, paid every month depending on how much it works. This guide explains how the second one is calculated, without prices that change every month, so you can estimate it with your own numbers.
Two costs: building and using
Building is the work of designing the task, connecting your tools, writing the instructions, testing with real cases and adjusting. It is quoted like any development: by scope.
Using is what the model provider (OpenAI, Anthropic, Google or another) charges every time the AI works. It is paid directly to the provider, with an account in your name, and it depends on volume: if the AI processes twice as many messages, it costs twice as much.
Paying usage directly to the provider has an upside: you see exactly how much is consumed, you set spending limits and you do not depend on a middleman reselling the service with a markup.
What a token is
Models do not charge by word or by message, but by tokens: small pieces of text the model reads and writes. A short word can be one token; a long one, two or three. As a rough reference, an English text uses somewhat more tokens than words.
They are charged separately:
- Input tokens: everything sent to the model: the instructions, the message or document to process and the business information it needs to answer.
- Output tokens: what the model writes. They usually cost several times more than input tokens.
One detail that surprises people: the instructions and reference documents are sent with every request. If every answer includes a five-page price list, those five pages are paid for every time.
How to estimate usage cost
With three numbers of your own and the provider’s pricing page you can do a reasonable calculation:
- Volume: how many times a month the AI works (messages answered, invoices read).
- Tokens per task: how much goes in and how much comes out each time.
- Price: what the provider charges per million input and output tokens.
An example with round numbers: 1,000 invoices a month. Each one sends about 3,000 tokens (the instructions plus the document) and returns about 500 (the extracted data).
| Item | Calculation | Per month |
|---|---|---|
| Input | 1,000 × 3,000 tokens | 3 million tokens |
| Output | 1,000 × 500 tokens | 0.5 million tokens |
| Cost | 3 × input price + 0.5 × output price | Depends on the model chosen |
With that calculation, comparing models is a matter of multiplying by another price. Do it with two or three: between a large model and a small one, the difference can be several times over.
Before putting anything to work, measure the real tokens with a trial: each provider’s console shows how much every request consumed. Estimates on paper tend to fall short.
The hidden costs
- Automation tools: if the flow runs on Zapier, Make or n8n in the cloud, they have their own subscription based on the number of runs.
- WhatsApp: the WhatsApp Business Platform charges for some types of messages, especially those the business starts.
- Server and storage: if the AI searches your documents, they have to be stored somewhere, prepared for search.
- Maintenance: checking how it answers, updating the information it uses and adjusting when a process changes or a new model comes out.
- Review time: what your team spends approving. It is a real cost, and it is the one that drops the most when the task is well built.
How to lower the cost without losing quality
- The right model for each step: classifying a message does not need the largest model. A small, fast one is usually enough, and it costs much less.
- Send only what is needed: instead of the whole catalog, the products related to the question.
- Rules where they are enough: what arrives structured does not need to go through a model (we explain it in AI or automation?).
- Short answers: ask for the data in a fixed format instead of paragraphs.
- Caching: several providers charge less for the part that repeats identically in every request, such as the instructions. Ordering what you send well lets you take advantage of it.
- Spending limits: provider consoles let you set caps and alerts. Set them up on day one.
What about a model on your own server?
Open models (such as Llama, Mistral, Qwen or Gemma) can run on your own server. There is no cost per token, but there is the server, which for large models needs powerful graphics cards, and the cost of maintaining it.
It usually makes sense when volume is very high and steady, or when the data cannot leave your infrastructure. For most businesses that are starting out, a provider’s API is cheaper and simpler. To compare options, read how to choose an AI model. And if you want an estimate with your numbers, tell us which task you have in mind.
Frequently asked questions
Why is AI usage paid separately?
Because it depends on volume and the model provider charges it. With an account in your name you see the real consumption, set limits and pay no middleman markup. Before starting, it is estimated with your numbers.
Can the cost run away?
It can, if nobody watches it: a bug that repeats requests or an unexpected volume. That is why it is worth setting spending caps and alerts in the provider account from day one, and checking consumption during the first weeks.
Is a model on my own server cheaper?
Only with high and steady volume. The server is paid for even when idle, and large models need expensive hardware. To start with, a provider’s API is almost always the better deal.
Do model prices change?
Yes, often. In recent years, for the same level of quality, they have tended to go down, and new models come out every few months. That is why your system should be able to switch models without being rebuilt.