When you read the word "unbelievable," you see one word. When an AI model reads it, it might see three pieces: "un," "believ," and "able." That gap — between how you chunk language and how a model chunks it — is the single most useful thing to understand about how these tools work. A token is the small unit of text an AI model actually reads and writes — usually a short word or a fragment of a longer one — and almost everything else, from cost to context limits to why the model sometimes miscounts letters, follows from it.
Let's build up to that from something you already know.
Start with how you break up text
When you skim a sentence, you don't process it letter by letter. You read in chunks — familiar words your eye grabs whole. A model does the same thing, except its chunks are fixed in advance by a process called tokenization: splitting text into the units the model was trained on.
Those units are tokens. A token is often a whole common word ("the," "note," "market"), but longer or rarer words get split into pieces, and spaces and punctuation count too. The model never sees "the cat sat" as a sentence of three words. It sees a list of tokens, each one converted into a number it can do math on.
A worked example
Take this short phrase:
JustJot makes note-taking effortless.
A typical tokenizer might split it like this (the ▁ marks a leading space):
Just · Jot · ▁makes · ▁note · - · taking · ▁effort · less · .
That's 9 tokens for 5 words. Notice the pattern:
- Common words ("makes") stay whole.
- A coined name ("JustJot") splits into familiar parts.
- A rarer word ("effortless") breaks into "effort" + "less."
- The hyphen and period are their own tokens.
A handy rule of thumb for English: 1 token ≈ 4 characters, or about ¾ of a word. So 1,000 tokens is roughly 750 words — about a page and a half. Other languages and code tokenize differently, often into more tokens per word.
Recap: A token is a chunk of text — whole word or fragment. Models read and write in tokens, not words.
Why this one idea explains so much
Once you can see tokens, three confusing things stop being mysterious.
1. Why AI tools charge the way they do. Most pay-as-you-go AI pricing is per token — counted on both what you send (input) and what you get back (output). A long pasted document costs more not because it's "harder," but because it's more tokens. Asking for a shorter answer genuinely costs less.
2. Why there's a limit to how much it can read. A model's context window is the maximum number of tokens it can hold in mind at once — your prompt and its reply combined. ("Context window" = the model's short-term memory, measured in tokens.) When a long conversation starts "forgetting" the beginning, you've filled the window and the oldest tokens have dropped off. It's a token budget, not a word count.
3. Why it sometimes flubs letters. Ask a model how many "r"s are in "strawberry" and it may stumble. That's because it never saw the individual letters — it saw the tokens "straw" and "berry." Counting characters means reasoning about something it doesn't directly perceive. Knowing this, you stop being surprised by it.
Where you'll meet tokens in real life
You don't have to do any of this math by hand, but you'll feel tokens everywhere once you know they're there:
- A "context limit reached" message means you've hit the token ceiling — trim or summarize, don't just rephrase.
- Pasting a 40-page PDF into a chat and getting a vague summary? It may have been truncated to fit the token window.
- Tight, well-structured notes don't just read better — they tokenize smaller, so an AI can fit more of them into one request.
That last point matters for anyone keeping a knowledge base. In JustJot.ai, AI features like semantic search and ask-your-notes work over your captured notes, and clean, focused notes give the model more room to work within the same token budget — better answers, less truncation.
Try this
Open any AI chat and paste two versions of the same request: one rambling, one tight. Ask each for a one-paragraph answer. You'll usually get a sharper reply from the tight version — partly because you spent fewer tokens saying it, leaving more room for the model to think. Seeing tokens, even indirectly, is the first step to using these tools deliberately instead of hopefully.