The article is assuming a little too much when it's claimed that the first implementation doesn't sort in place.

It's not clear how GHC would optimize this (even more if you don't use the unsorted list for anything else), and it would be interesting to see the benchmarks for that version too.