One "interesting" thing about US AI companies is that they do not really show what the models reason about (i.e., how they work). They literally have an LLM in front of their LLM, summarizing what's happening in the back.
This is rather obvious with the Google / OpenAI "title + paragraph of text" way of things, but if you look closely at Anthropic models output, they actually do the same, but at the sentence level. You get a burst of text, then pause, then burst of text, then pause.