At that point it would be useful. Although, I wonder if it wouldn’t make more sense to train one specifically for the job. Current LLMs can predict HTML, sure, but they’re large and slow for the task.