More tokens = fewer options for the final string. It's not any more complicated than that and it doesn't require any reasoning, just an autoregressively trained statistical model of text, but no, it has to be "the model thinks harder if it outputs more tokens".