Whenever someone mentions the price or memory of an AI, one word pops up: “tokens”. AI tokens are not the ones for the bumper cars. Although the idea isn’t that far off: you pay per ride.
AI tokens in one sentence
A language model doesn’t read whole words. It chops text into pieces called tokens. A token can be a short word, part of a long word, or a comma with an inflated ego.
Why tokens count (literally)
- Price: most AI services bill by the number of tokens sent and received. Chattiness has a cost.
- Memory: each model handles only so many tokens at once, its “context window”. Beyond that it forgets the beginning, like us at the end of a three-hour meeting.
- Speed: the more tokens to write, the slower the answer. A novel takes longer than a text message.
AI tokens in a small business: the 80-page contract
You want an AI to answer questions about your contracts. Sending all 80 pages with every question is like rereading the whole dictionary to check one spelling. Better to give it only the relevant passages: that’s RAG, search across your documents.
A handy habit: before any AI project, take ten real documents, count the tokens they produce with your provider’s tools, and multiply by your monthly volume. One hour of work, and you won’t discover the running costs after launch.
The limits (this is where we stop joking)
Tokens per word vary by language and model: estimate costs on a real sample of your texts, not a wild guess.
At Octopote we measure that cost from the prototype onwards, so no surprises. To estimate a project without counting tokens by hand, try our quote simulator. And to pick the right AI, read machine learning vs generative AI.
