Send the same prompt to the same model twice — once with thinking off, once with thinking on — and watch both transcripts side by side.