My understanding is that it mainly fails when you try it in speech mode, because it is the fastest model usually. I tried yesterday all major providers and they were all correct when I typed my question.