It's not a "benefit", it's simply how the technology works - the underlying model just fundamentally works on tokens as it's atomic inputs.

The models don't know anything about words, just tokens.