You mention thinking tokens as a side note, but their existence invalidates your whole point. Virtually all modern LLMs use thinking tokens.