Tokenizer fertility
How many tokens a tokenizer needs, on average, for one word of a given language. Lower is cheaper and faster.
Also: fertility, token tax
Fertility measures how efficiently a tokenizer covers a language. English text in a modern tokenizer averages a little over one token per word. Hindi, Bengali, Tamil and other Indian languages have often needed three or more tokens per word in tokenizers built mainly for English, so the same message costs more, fills the context window sooner and is generated more slowly. Indian model builders such as Sarvam AI have trained their own tokenizers to bring fertility down, and newer global models have enlarged their vocabularies for the same reason.