--- title: What is a token? url: https://stowly.dev/learn/what-is-a-token description: Subword chunks, per-million billing, input vs output cost. docs_index: https://stowly.dev/llms.txt version: v0.1 lastUpdated: 2026-09-17 --- > Fetch the complete documentation index at https://stowly.dev/llms.txt before exploring further. Models don't read words. They read **tokens** — short chunks of text that get turned into numbers before anything else happens. A token is often a whole common word (`the`), sometimes part of a word (`un` + `happiness`), sometimes a single character or symbol. As a rough rule, **one token is about three quarters of an English word**, or around four characters. ## Why chunks instead of words? Splitting on spaces feels obvious until a model meets a word it has never seen — a new product name, a typo, another language. Whole-word vocabularies would need millions of entries and still hit unknown words constantly. Modern models split text into **subwords** instead. Common words keep their own token; rare words fall back to familiar pieces. The model never gets stuck: anything can be spelled out from parts it already knows. Type anything above. The colors are approximate — every model slices text slightly differently — but the count and the cost use the same math as the calculator. ## Why tokens decide your bill Providers charge **per million tokens**, with input and output priced separately. Output almost always costs more: looking up an input token is cheap, generating each output token means running the whole model again. Think of it like a factory — unloading raw materials is cheap, manufacturing each new item is not. So two things decide every invoice: **how many tokens your text becomes**, and **each token's price**. Change either and the number moves. ## Quirks worth knowing Because models see chunks and not letters, they can be oddly bad at letter-level questions — like counting the r's in "strawberry". They never see the letters, only the chunks. Emoji, code, and non-English text also split less efficiently, which means they cost more per idea than plain English. ## What to do with this - Keep prompts tight: every token is billed. - Prefer one precise word over three vague ones. - Price the prompt before shipping it — [open the calculator](/ai-token-cost-calculator) and paste the exact text your app sends. - Check where each rate came from on [Sources](/sources), and how the math works in [Methodology](/methodology).