AI doesn't read words — it reads tokens. They're why API pricing, context limits, and speed are all measured the way they are. Here's what a token actually is, in plain English.
A token is a chunk of text that an AI model processes as a single unit — and it's usually not a whole word. On average, a token is about ¾ of a word (roughly 4 characters in English), so common words are one token but longer or rarer words split into several. AI models don't read letters or words the way we do; they read sequences of tokens. This matters because tokens are the unit behind almost everything you deal with: API pricing is per token, a model's [context window](/what-is-a-context-window-explained-2026) is measured in tokens, and generation speed is 'tokens per second.' Understand tokens and a lot of AI suddenly makes sense.
How tokenization works
Before a model sees your text, a tokenizer breaks it into tokens. Common words like 'the' or 'running' become a single token; a rarer or longer word like 'tokenization' might split into 'token' + 'ization', and unusual strings (code, odd names, other languages) split into more pieces. The model then works entirely with these tokens — predicting the next token, over and over, to generate a response. That's literally all a language model does at inference: guess the next token, add it, repeat. Because everything is counted in tokens, the practical conversions are worth memorizing: 1 token ≈ 4 characters ≈ ¾ of a word, so 1,000 tokens ≈ 750 words, and a typical page of text is roughly 500 tokens. When a model has a '128K context', that's ~96,000 words it can hold at once; when a cloud API charges '$1 per million input tokens', that's about $1 per 750,000 words you send it.
AI reads text as tokens — chunks of about ¾ of a word — which is why cost, context, and speed are all measured that way. · Unsplash
Why tokens matter for local AI
When you run AI locally, tokens show up in two places you'll actually feel. First, speed: local performance is measured in tokens per second, and it's the number to watch — a good GPU might generate 40-100 tok/s on an 8B model (fast enough to read comfortably), while an overloaded setup crawls at a few tok/s. Second, context and memory: since the context window is measured in tokens and a longer context uses more VRAM, tokens are the link between 'how much text I'm feeding the model' and 'will it fit on my card.' The nice thing about running locally is that, unlike cloud APIs, you don't pay per token — once the hardware is yours, tokens are free, so the only limits are speed and context, not cost. That's a big part of why running AI locally appeals to heavy users.
Quick answers
What is a token in AI?
A token is a chunk of text that an AI model processes as a single unit — typically about three-quarters of a word, or roughly 4 characters in English. AI models read and generate text as sequences of tokens rather than whole words or individual letters. Common words are usually one token, while longer or rarer words split into several. Tokens are the fundamental unit behind AI: pricing is per token, context windows are measured in tokens, and generation speed is measured in tokens per second.
How many words is 1,000 tokens?
About 750 words. The general rule is that one token equals roughly three-quarters of a word (or about 4 characters in English), so 1,000 tokens is approximately 750 words, or about 1.5 pages of text. This conversion helps you interpret AI specs: a 128K-token context window holds roughly 96,000 words, and a cloud API charging $1 per million input tokens costs about $1 per 750,000 words of input. Code and non-English text tend to use more tokens per word.
Why is AI measured in tokens instead of words?
Because tokens are the actual unit AI models work with internally. A tokenizer breaks text into tokens before the model processes it, and the model generates responses one token at a time by predicting the next token. Since the model's real operations happen in tokens, everything built around it — pricing, context limits, and speed — is measured in tokens for accuracy. Words are ambiguous (lengths vary, languages differ), while tokens are the precise, consistent unit the model computes on.
