I would say that it's very strong evidence that it is thinking, if that "thinking out loud" output affects outputs in ways that are consistent with logical reasoning based on the former. Which is easy to test by editing the outputs before they're submitted back to the model to see how it changes its behavior.