Subword tokenization keeps frequent words as single tokens while splitting rare or unfamiliar words into recognizable pieces. This is why a common word like 'cat' may be one token, while an unusual word like 'unbelievable' can break into several subword tokens such as 'un', 'believ', and 'able'. The model can then handle words it has never seen as a whole.
How ChatGPT Writes One Token at a Time