Token

AI doesn't read and write text character by character — it works in "tokens." A token is a frequently used chunk of text. One English word is usually 1-2 tokens, while for Korean, roughly one character tends to be 1-2 tokens. A rough rule of thumb: "1 token ≈ a few short syllables."

This unit matters because money and limits are all calculated in tokens. AI companies list pricing as "dollars per million tokens," and free usage caps and the amount of text you can process at once (the context window) are also measured in tokens.

Just as your water bill is calculated by the "liter," your AI bill is calculated by the "token" — remember that, and pricing articles suddenly make sense.

"The new model's API price is $3 per million input tokens" — meaning it costs just a few dollars to process hundreds of pages of text. Note that input (what you feed in) and output (what the AI writes) are priced differently.

1. Search "OpenAI tokenizer" to open the tokenizer demo page.
2. Enter an English sentence and a Korean sentence with the same meaning, then compare the token counts.
3. Korean usually gets split into more tokens — meaning the same content costs more in billing units. This is exactly where the "Korean AI is more expensive" debate comes from.

Both are numbers used to describe scale, so when an article mentions something like '70 billion,' it's easy to lose track of which one is being discussed. One is the amount of text you send and receive, the other is how big the model itself is. If it grows with how much you use it, it's the former; if it's fixed when the model is built and never changes, it's the latter.