Why may a short Chinese phrase and a short English word have unrelated token counts?

Answer: Token chunks differ by language

Only spaces decide it — Spaces matter, but they are only part of the story. Tokenizers learn chunks from training text, so Chinese characters, phrases, punctuation, and English subwords can be grouped by different statistical habits. Space is a visible clue, not the whole mechanism.

Token chunks differ by language ✓ — Token chunks are learned from text, and languages distribute characters and word boundaries differently. A tokenizer may treat a common Chinese phrase as one chunk while splitting a rare English word, or do the reverse in another encoding. This is why token budgets are language-dependent rather than just length-dependent.

Characters equal tokens — A character is not guaranteed to equal one token. Some phrases become compact learned chunks, while some rare words or symbols split into smaller pieces. The unit the model sees is learned from text statistics, not copied from a writing system's visible characters.