What are AI tokens? A plain-English guide to how AI pricing actually works.

§ 01 — What is a token, and why does AI pricing use it instead of something simpler?

Text gets broken into tokens before an AI model processes it, because models work on numbers, not letters, and tokens are the numeric building blocks. As a rough rule of thumb, 750 words is close to 1,000 tokens, though the exact ratio moves around depending on the language and how unusual the words are.

Pricing splits into two kinds. Input tokens are everything you send the model: your question, any document you've pasted in, and (in a back-and-forth chat) usually the earlier messages too, sent again each time. Output tokens are what the model sends back. Providers typically charge more per output token than per input token, since generating text costs more computing than reading it.

This is why a short question about a long document isn't cheap: the document is the input, and you're paying for all of it even though your question was one line.

§ 02 — What does this actually cost in practice? A worked example.

Take a business summarising a 20-page report every working day. Twenty pages is roughly 8,000–10,000 words, call it 12,000 tokens once formatting and structure are accounted for. A useful summary back might run 300–500 tokens.

Multiply that by a working month (call it 22 days) and you're looking at somewhere in the order of 250,000–280,000 input tokens and 7,000–11,000 output tokens a month for this one task, comfortably under half a million tokens in total.

Model pricing varies enormously by tier, and published rates change often enough that quoting a precise figure here would be wrong within months. As a rough sense of scale: the cheapest general-purpose models on the market typically price at a fraction of a cent per thousand tokens, while the most capable frontier models can charge ten to a hundred times more per token for meaningfully better reasoning on harder tasks. For a task as straightforward as summarising a structured report, that gap matters, because the cheaper tier is usually good enough, and running this specific example through it typically lands in the low tens of dollars a month, not hundreds. The same task pushed through a top-tier frontier model, for no real accuracy benefit on a simple summarisation job, could run several times higher for identical output quality on this kind of task.

That gap, choosing the right-sized model for the job rather than the most impressive one, is most of what separates a well-scoped AI build from an expensive one, and it's exactly the kind of detail that should show up as an honest cost in a proposal rather than a surprise on the first invoice.

§ 03 — When does token cost stop mattering, and what should you watch for instead?

For most small-business automations, the honest answer is: almost immediately. A task like the reporting example above, run daily, typically lands in the tens of dollars a month in raw compute. Compare that to the cost of properly building, testing, and handing over the automation, which is normally the far larger number, and token cost turns out to be closer to a rounding error than a genuine budget line.

The caveat runs the other way too: it's worth being wary of anyone pricing a project as though compute is the expensive part, especially for tasks like the kind described in the five-question test for whether something's worth automating. If a quote leans heavily on "ongoing AI usage fees" for a task that, by the maths above, should cost tens of dollars a month to run, ask directly what that fee is actually covering.

Where token cost does start to matter is at genuine scale, processing thousands of documents a day, or running very long documents through very capable models repeatedly. At that point the model choice and the token math become a real design decision, not a rounding error, and it's worth getting specific advice rather than working from rules of thumb.

If you're being quoted on an AI project and the pricing doesn't add up against the maths here, send the brief and I'll take an honest look at it.