It’s better for them

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

The most important feature of a tokenizer is dividing the inputs into independent values that the neural network can work with. It's not the size.