Are you sure? It would be great to get official/semi-official validation that thinking is or is not resolved to a token embedding value in the context.

You can read the model cards. Claude thinks in regular text, but the summarizer is to hide its tool use and other things (web searches, coding).