A similar question to this was asked in 2016.

The fivethirtyeight.com team, lead by Nate Silver, was able to predict exactly the 2012 state-by-state electoral map for the presidential election. In 2008 their prediction missed on only a single state, Indiana, which Obama won by a 0.1% margin. However, in 2016, they missed several states -- such as Michigan, Pennsylvania, and Wisconcin -- ultimately missing the final election outcome.

See for example this writeup on Mashable after his 2012 predictions aced the electoral map.

Will the 538 model repeat its 2012 success and assign greater than 50% probability to the correct outcome in every state and electoral district?

This will resolve as positive if the final prediction from fivethirtyeight.com before the election matches exactly the electoral results post-election for each state. In other words FiveThirtyEight must predict the modal outcome for the election as being the exact real outcome that actually happened. (Maine and Nebraska allow their electoral votes to be split; for these two states, success requires all the districts to be called correctly.) If the site has multiple prediction models showing as of the day before the election, whichever is highlighted as their best pick will be used.