Do the thinking models actually think?