If you were to keep a single technical term from this whole glossary in order to control your costs, this would be it. The token is the unit in which a model reads, writes and bills. Everything else follows from it.
It is also the word that explains surprise invoices. An automation that seemed to cost a few cents ends up at several hundred euros a month, and the answer is almost always in the number of tokens consumed without anyone noticing.
Definition
A token is a piece of text as the model cuts it up. It is neither a character nor exactly a word: it is an intermediate unit, often a whole short word or a fragment of a long one.
In English, count roughly four tokens for every three words. A 300-word email therefore weighs around 400 tokens, an A4 page about 700, a twenty-page contract around 14,000.
Languages are not equal at the cutting stage. English uses fewer tokens than French for the same content, because models were trained mostly on English. For identical content, an instruction written in English often costs 25 to 30 percent less.
Why this unit decides your bill
Every provider bills by the token, with a distinction many discover too late: input and output are not priced the same.
| Direction | What it is | Relative price |
|---|---|---|
| Input | Everything you send: instruction, history, documents | The cheaper side |
| Output | What the model produces | Often three to five times more |
That asymmetry has a counter-intuitive consequence. Asking for a ten-line summary of a fifty-page document costs little, while asking for a long text from a one-line instruction costs a lot. It is not the question that weighs, it is the answer.
The three traps that blow up the bill
History resent with every message
In a conversation, the entire previous exchange is resent to the model with each new message. By the twentieth message you are paying nineteen times over for the accumulated context. It is the biggest source of waste, and the least visible.
The whole document pasted in
Sending a hundred-page manual to ask about one paragraph works, but costs a hundred pages per question. That is precisely the problem RAG solves.
The loop with no ceiling
An AI agent chaining steps consumes on every pass. With no limit on the number of steps, an agent stuck on an unsolvable problem can burn a monthly budget overnight.
The right monitoring habit is not watching the total at month end, but the number of tokens per operation. A total that doubles because activity doubled is healthy. A total that doubles at constant activity signals a leak.
How to bring consumption down
Start a fresh conversation. As soon as an exchange drifts from its original subject, open another one with a short summary. That cuts accumulation at the root.
Use caching where it exists. If the same long instruction is sent on every call, most providers let you cache it at a reduced rate. At volume, the saving is substantial.
Match model size to the task. A small model often costs ten times less than a large one. On sorting or extraction, quality is equivalent.
Ask for short answers. Since output is the expensive side, adding "in three sentences" acts directly on the bill.
How many tokens are in your usual content
These orders of magnitude are enough to size a project without a counting tool. They apply to English; French is less economical, closer to five tokens for every three words.
| Content | Words | Tokens, roughly |
|---|---|---|
| A short message | 30 | 40 |
| An everyday email | 150 | 200 |
| An A4 page | 500 | 670 |
| A detailed product description | 800 | 1,070 |
| A blog article | 1,500 | 2,000 |
| A twenty-page contract | 10,000 | 13,300 |
| A two-hundred-page book | 80,000 | 107,000 |
Two reference points follow from that table. A thirty-message conversation carries, by the last message, the equivalent of a blog article as input, and you pay for it again with every new message. And a document you paste to ask a single question costs its full weight on every question asked, which becomes absurd beyond three or four.
An estimate, not an exact count. Every model has its own splitting: this simulator reproduces the principle and gives the right order of magnitude, not your provider's figure.
Frequently asked questions
How do I know how many tokens my text contains?
Most providers offer a counting tool, and their professional interfaces display the real consumption of each call. For a quick estimate, the three-tokens-per-four-words rule is more than enough to size a project.
Does a token cost the same everywhere?
No, the gap between models runs from one to fifty. This is why thinking in tokens rather than in euros makes comparison possible: measure how many tokens your use consumes, then apply the rate of the model you are considering.
Does the length of my question affect the quality of the answer?
Up to a point, yes: giving context clearly improves the result. Beyond that, the effect reverses. An instruction buried in ten pages of incidental material is followed less well than a clean instruction with two well-chosen extracts.
How do you keep these costs under control in an automation?
By setting a step ceiling and a budget alert from the start, before the automation ever runs in production. Our n8n course covers that alongside building the scenarios themselves, rather than as an afterthought.