That may have always been the case. After all, humans press the button while they continue thinking, and even take some time to think while they start to answer. It's only logical for a computer of that scale to press the button even when it did not fully finish computing, winning a second or two, let alone milliseconds.