Thinking models think through talking, don't reveal that talking, then answer by again thinking through talking. It's kinda funny in a way.

I always expose reasoning traces. How else can you seriously debug?