Haskell is also astoundingly terse. In Java and C my data type declarations were too long to fit on a page, full of redundancies and boilerplate. In Haskell if you want to make, say, a data type that is either an X or an O (suppose you're writing tic-tac-toe), you could do it in four words: 'data XO = X | O'. (Notice that there's not even a natural way to do that in Java or C, because they don't have sum types; you'd have to make a type that has a flag to indicate whether it is an X or an O. That gets really complicated if they're supposed to have different data associated with it -- but in Haskell, if the X is supposed to carry a float and the O is supposed to carry a string, you just add two more words.)
Pattern matching also helps with terseness. I don't have time even to write the last paragraph so I'll skip this one.
Purity keeps you from tripping up on IO-related errors. It lets you be much more certain that things are working. It also forces you to keep the IO in a thin top-level layer of your program, which might sound like a pain but once it feels natural you'll find yourself moving faster than you could before.
To be sure, Haskell has features that I don't use. But purity, sum types, pattern matching, and the unusually rigorous (it's complete!) type system are all critical to its value to me.