Remember that we're only evaluating the models against one single prediction (the election itself). Because we now know that the election was very close, we also know that either candidate probably had a reasonably strong chance at winning. That means that for the 98% model to be accurate, voter turnout on election day would have had to have been an astoundingly improbable statistical aberration -- as in, likely tens or hundreds of thousands of voters would have had to have behaved differently than the data would have suggested. On the other hand, the prediction of a 75% chance of victory is also a prediction of a fairly close race. So, 538's model was more consistent with the results, by far.
As far as when that transition would occur, I'm not sure that can be known. We'd need to be training these statistical models against far more predictions. My only point here is that not all the predictions were off base, and the talking point that all the predictions were wrong is a lie. 538 didn't predict that Hillary would win, they predicted she had a 75% chance at winning.