I read somewhere that the test might switch models mid conversation which stops the other from reading the previous model's reasoning. I don't know if that's true, but it sure feels like that quite often in the last days.