A token is the little chunk of text an AI model actually reads and writes — not a whole word, not a single letter, but a piece in between. Every time you use an AI model, your text gets sliced into tokens going in, and the model produces tokens coming out. That one idea quietly explains three things people find confusing: why AI bills are shaped the way they are, why there’s a limit to how much you can paste in, and why “just make it cheaper” isn’t simple. Let’s make it click.
Imagine text as Lego. A model doesn’t see sentences; it sees a stream of standard bricks it snaps together. Each brick is a token. Common short words like “the” or “cat” are usually one brick. Longer or rarer words get built from several — “tokenization” might be three or four bricks. As a rough rule of thumb in English, one token is about four characters, and 100 tokens is roughly 75 words. So a page of text is a few hundred bricks, and a long document is thousands.
Why not just use words? Because models need a fixed, manageable vocabulary of pieces that can spell anything — names, typos, code, emoji, other languages — without a dictionary that’s infinitely long. Sub-word tokens are the compromise: a few tens of thousands of bricks that can assemble any text you throw at them. It’s the same reason a small set of Lego shapes can build almost any model — flexibility from a limited, standard kit.
The mental model: tokens are the currency of AI. The model reads them, writes them, and — on cloud services — charges for them. Understand tokens and the pricing page suddenly makes sense.
Cloud AI is billed per token — so much per thousand tokens in, so much per thousand out. That’s why a long prompt with a big pasted document costs more than a short question, and why an assistant that re-reads a huge conversation every turn gets expensive fast. It’s also why “summarize this 80-page PDF” can quietly be a large bill: those pages are tens of thousands of tokens the model has to ingest before it says a word. Once you think in tokens, you can predict cost instead of being surprised by it.
There’s also a hard ceiling on how many tokens a model can hold in mind at once — its context window. Everything has to fit: your instructions, the document, the back-and-forth, and the answer, all counted in tokens. When people say a model has a “128K context,” they mean 128,000 tokens of working memory. Blow past it and the model forgets the earliest parts. We unpack that limit fully in what is a context window — tokens are the unit that window is measured in.
Here’s the part that matters strategically. On a cloud service, tokens are a meter — every brick in and out is a charge, forever. When you run a model on your own machine, tokens still exist (the model still chops text into them), but nobody’s charging per brick. Your cost is the hardware and the electricity, flat, whether you process a hundred tokens or a hundred million. For anyone doing heavy work — long documents, big context, constant use — that’s the whole argument for local AI in one number. It pairs directly with how you’d then customize a model, which we cover in what is fine-tuning.
A token is the sub-word chunk an AI model reads and writes — roughly four characters, about 75 words per 100 tokens. Tokens set your bill on cloud services and define how much a model can hold at once. Grasp them and AI pricing, context limits, and the case for running models on your own hardware all stop being mysterious — they’re all just tokens, counted.
Stop paying by the brick. ABUZ8 OS runs models on your own machine — tokens still work, but there’s no per-token meter. It’s in early access (live and hardening, no card). Explore the free tools or join early access.