Every tokenizer, ranked on any test set

UTF-8 bytes per token on held-out text; higher means fewer tokens for the same text. Filter by provider or vocabulary size, or switch from single lines to whole documents.