How does a sentence become tokens?
Before a model sees your sentence, a small program called a tokeniser cuts it into pieces from a fixed vocabulary and swaps each piece for a number. The model never sees letters, only the list of numbers. OpenAI's newer tokenisers have a vocabulary of about 200,000 pieces, and the rule of thumb of four characters a token comes from OpenAI's tokenizer documentation.
The vocabulary is built by a method called byte pair encoding, adapted for language models by Rico Sennrich and colleagues at the University of Edinburgh in 2016 from an older data-compression trick. Start with single characters. Count which pair of neighbours turns up most often in a huge sample of text, glue that pair into a new piece, and repeat, tens of thousands of times. Frequent words end up as one piece. Rare words never get glued all the way, so they stay as several.
Why is one word sometimes three tokens?
Because the vocabulary follows how often text appears, and meaning has nothing to do with it. "Book" and "cruise" are common enough to be single pieces. Woolloomooloo is rare in the text the tokeniser was built from, so it gets spelled out in chunks, and an Australian spelling can cost a piece more than the American one in a tokeniser built mostly from American text. Numbers are often split every few digits, a typo breaks a word apart, and the space before a word usually travels with it as part of the token. OpenAI's guidance on tokens also notes that other languages generally take more tokens for the same meaning.
It is also why an AI can trip on a question about letters, such as how many r's are in "strawberry". It never saw the letters. It saw two or three pieces.
Why are AI services priced per token?
Because every token costs computing. To write each token of a reply, the model runs a full pass through its billions of internal values, so a 400-token answer is 400 passes. Reading your input costs less per token because it can be processed in one go, which is why the price lists from OpenAI, Anthropic and Google quote input and output separately, per million tokens, with output the dearer of the two. The full story is under inference.
The other limit counted in tokens is the context window: the most text a model can hold at once, from your instructions through to its reply.
How many tokens does a vet clinic's website chat get through in a month? (example figures)
A vet clinic in Wagga Wagga runs an AI chat on its website for 400 conversations a month. Every turn of a chat re-sends the clinic's instructions and price list (1,500 tokens) plus the conversation so far, because the model keeps nothing between turns. An average chat has 6 turns, with 30-token questions and 80-token replies. The rates are example figures in Australian dollars.
| Where the tokens go | Sum | Result |
|---|---|---|
| Clinic instructions, re-sent each turn | 1,500 × 6 turns | 9,000 tokens a chat |
| Questions and earlier turns, re-sent | Grows by 110 tokens a turn | 1,830 tokens a chat |
| Replies written | 80 × 6 turns | 480 tokens a chat |
| Input for the month | 4.33 million tokens at $4.50 a million | $19.49 |
| Output for the month | 192,000 tokens at $22.50 a million | $4.32 |
| Total | $23.81 |
More than 80 percent of the input is the same 1,500 tokens of clinic information, sent again on every turn. Trimming it to 600 tokens brings the month down to about $14. OpenAI and Anthropic both also offer prompt caching, a lower rate for input that repeats.
What reaches the model when you talk to Cora?
Tokens made from your words: in a conversation with CoraVision, speech becomes tokens like any other input, and a model can only work with what those tokens contain. So rather than relying on the figures you remember, CoraVision reads your accounts. It reads Google Ads, Meta and Google Analytics 4 as soon as they are connected and again each night, and during a session you start, Cora sees your screen as still images that are not saved to your account.
Questions
How many tokens is a page of writing?
Using OpenAI's rule of thumb that 100 tokens is about 75 words, a 500-word page of plain English is roughly 670 tokens. Tables, code, unusual names and other languages run higher. Pasting a sample into a tokeniser gives the real count for that model.
Do pictures and voice use tokens too?
Yes. Models that take images cut them into patches and count those as tokens, and the model makers publish how a picture's size converts into a token count. Speech is sliced into short stretches of audio that become tokens in the same way.
Where can I see how my own text splits?
OpenAI publishes a free tokeniser page at platform.openai.com/tokenizer. Paste in any text and each token is shaded separately, with a count. Other model makers split text a little differently, so treat it as a close guide rather than an exact one.