I don't understand why your conclusion is that "the model must be thinking beyond the next token": the model doesn't need to do that to generate a well-formed sentence because it's not constrained by the size of the sentence.