What is a token in AI chat?
Updated
A token is the unit of text a language model actually reads and writes: often a whole short word, sometimes part of a longer word, a space plus a word, or a punctuation mark. In English, one token averages roughly three quarters of a word. Context windows, reply length limits and API usage are all counted in tokens, not words or characters.
Try it now, free, no account neededStart a story with Lady KohanaA fox spirit wants you as her muse. And her editor.How text becomes tokens
Before a model sees your message, a tokenizer splits it into pieces from a fixed vocabulary. Common words are usually one token. Rare words, names and invented terms get split into several. A sentence like:
*Kohana dips her brush in the ink and frowns at your manuscript.*
is about 15 words but may be closer to 20 tokens, because the asterisks, the name "Kohana" and "manuscript" may each take more than one piece. Every model family has its own tokenizer, so the same text can be a slightly different number of tokens on Gemma than on another model.
Why tokens matter in roleplay
- Context: the context window is a token budget. UnsavedAI models allow 61,440 tokens including the reply.
- Reply length: settings like max tokens or Max Response Length cap how long a reply can be. Set it too low and replies stop mid-sentence; see why replies cut off.
- Card size: a long character card costs tokens on every single reply, because it is sent every time.
- Speed: models write one token at a time, so a 600-token reply takes longer than a 200-token one.
Quick rules of thumb
These are approximations for English prose, useful for planning rather than exact counts:
- 100 tokens is about 75 words, a short paragraph.
- A typical roleplay reply is 150 to 400 tokens.
- A solid character card is often 500 to 1,500 tokens.
- Other languages, code and heavy formatting usually take more tokens per word.
SillyTavern shows a token count for each card field and for the whole prompt, which is the easiest way to see where your budget goes.
Common mistakes
- Treating word limits and token limits as the same. A 300-token reply is closer to 225 words.
- Padding cards with decorative symbols, long separators or repeated headings. They cost tokens and teach the model nothing.
- Assuming every app counts the same way. Each one formats the prompt differently, and that formatting adds tokens too.
Tokens in UnsavedAI and other apps
On the UnsavedAI site you do not have to count anything: the chat trims old history automatically. If you use an API key in SillyTavern or JanitorAI, you set context size and max reply length in tokens yourself; the JanitorAI setup guide lists safe values. Plans on UnsavedAI are not priced per token. To see how a long, descriptive reply feels, try Lady Kohana, who writes like the novelist-muse she is.
Questions people ask
How many tokens is a word?
In English, about 1.3 tokens per word on average. Short common words are usually one token; long or unusual words are several.
Do emojis and asterisks use tokens?
Yes. Every character of text, including formatting marks and emojis, ends up in some token.
