I mean it might lead to better performance on the model side. So the tokenizer is better but more expensive.