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

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

AI models work on numbers, not letters, so text gets broken into tokens before the model processes it. Tokens are the numeric building blocks. As a rule of thumb, 750 words comes to about 1,000 tokens, though the exact ratio shifts with the language and how unusual the words are.

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

That's why a short question about a long document isn't cheap. The document counts as input, so you pay for all of it even though your question was one line.

What does this cost in practice?

Say a business summarises a 20-page report every working day. Twenty pages is roughly 8,000 to 10,000 words, so call it 12,000 tokens once you count formatting and structure. A useful summary might be 300 to 500 tokens.

Over a working month of about 22 days, that's somewhere around 250,000 to 280,000 input tokens and 7,000 to 11,000 output tokens for this one task. Comfortably under half a million tokens in total.

I won't quote exact prices, because published rates change often enough that any figure here would be wrong within months. To give a sense of scale, though: the cheapest general-purpose models usually cost a fraction of a cent per thousand tokens, while the most capable frontier models can charge ten to a hundred times more for better reasoning on hard problems.

For summarising a structured report, the cheaper tier is usually good enough. Run this example through it and you'll typically land in the low tens of dollars a month, not hundreds. Push the same job through a top-tier model and it could cost several times more for output that's no better.

Picking the right-sized model for the job, rather than the most impressive one, accounts for most of the difference between a well-scoped AI build and an expensive one. It's also the kind of detail that should appear as an honest line in a proposal, not as a surprise on the first invoice.

When does token cost stop mattering, and what should you watch instead?

For most small-business automations, almost immediately. A daily task like the reporting example usually costs tens of dollars a month in raw compute. Building, testing and handing over the automation properly is normally the much bigger number, which makes the token bill closer to a rounding error than a real budget line.

That cuts both ways. Be wary of anyone pricing a project as if compute is the expensive part, especially for the kinds of tasks covered in the five-question test for whether something's worth automating. If a quote leans heavily on "ongoing AI usage fees" for a job that should cost tens of dollars a month to run, ask what that fee actually covers.

Token cost starts to matter at real scale: thousands of documents a day, or very long documents run repeatedly through very capable models. At that point the choice of model and the token maths become a genuine design decision, and specific advice is worth more than rules of thumb.

If you've been quoted on an AI project and the pricing doesn't add up against the numbers here, send me the brief and I'll give you an honest read on it.