Why can one emoji cost an AI language model several tokens when it takes up a single character on your screen?

Answer: The tokenizer works on the text's bytes, and one emoji is stored as four or more bytes

The tokenizer stores every character as one token but adds a separate token for each colour in the emoji — Wrong. A tokenizer never sees colours; an emoji's picture lives in your device's font, and the model receives only the code that names it. That code is what costs several bytes, and therefore sometimes several tokens.

The tokenizer splits text at spaces, and an emoji with no space before it merges into the word next to it — Wrong. Modern tokenizers do not simply cut at spaces; they learn frequent chunks of bytes, and an emoji is handled as its own byte sequence whether or not a space sits beside it. The extra tokens come from the emoji's own length in bytes, not from gluing it to a neighbour.

The tokenizer works on the text's bytes, and one emoji is stored as four or more bytes ✓ — Correct! Text reaches a byte-level tokenizer as UTF-8 bytes, not as the characters you see, and a single emoji is encoded as four bytes, more when it carries a skin tone or is built from several symbols joined together. The tokenizer then merges byte sequences it has learned: common emoji may collapse into one token, rarer ones stay as several pieces. A tiny picture can therefore be computationally larger than it looks.