No, tokenization is not the only reason. A next-word predictor has fundamentally a hard time executing algorithms, even as simple as counting.