Fertility (tokenization)

Fertility measures how many tokens a tokenizer generates per word of input text. A fertility of 1.0 means each word maps to a single token; higher fertility means words are split into more pieces. Languages underrepresented in the tokenizer's training data tend to have much higher fertility, which directly increases cost (since pricing is per-token) and degrades model accuracy. Research shows that transformer attention scaling means a 2x increase in fertility can quadruple computational cost.