What are tokens? (and why your AI bill is measured in them)

By Dylan Bailey, Certified Claude Expert · Updated 4 August 2026 · 5 min read

Tokens are one of those terms you can happily ignore right up until you're looking at a bill, or wondering why a long document made an AI tool slow down and start losing the thread. Both come back to the same unit.

What a token is

Models don't read letters or words. They read tokens — chunks that are usually a whole common word, or a piece of a longer one.

"The invoice is overdue" is about five tokens. A rarer word like "reconciliation" might split into three or four on its own. Punctuation, spaces and line breaks all count.

The useful conversions:

• 1 token ≈ 0.75 English words
• 1,000 tokens ≈ 750 words ≈ 1.5 pages
• A typical email ≈ 200–400 tokens
• A 20-page report ≈ 10,000 tokens

Why tokens show up twice

In what you pay. Usage-based pricing counts tokens in and tokens out, with output typically costing more than input.

In what fits. The context window — everything the model can consider at once — is measured in tokens too. Fill it and the earliest material drops out of view.

So a long attachment costs you twice: money, and room. It's the same constraint wearing two hats.

An Australian small-business example

A Newcastle logistics firm sets up an assistant to answer questions about their operations manual. To be safe, they attach the entire 180-page manual to every conversation.

Each question now sends roughly 90,000 tokens before anyone types anything. The answers are slower, more expensive, and — counter-intuitively — less accurate, because the relevant paragraph is competing with 179 pages of material that isn't.

Splitting the manual by section and attaching only the relevant one cuts the cost dramatically and improves the answers. Less really is more here.

When to care, and when not to

For a team of ten using AI through normal subscriptions, tokens are almost irrelevant to your budget. The subscription covers it, and any time saved dwarfs the difference.

Tokens start mattering when:

• You're running something automated across thousands of records
• You're on usage-based pricing rather than a flat plan
• You're attaching large documents repeatedly
• Answers are degrading on long documents — a context problem showing up as a quality problem

The practical takeaway

Don't optimise tokens for their own sake. Do develop one habit: send the relevant part, not everything you have.

That single habit reduces cost, improves accuracy and avoids context problems simultaneously — which is unusual, because those three normally trade against each other.