I write a lot of little awk scripts, but if they grow past ~5 lines, they usually get rewritten in Lua. (Perhaps eventually with inner loops in C.) Still, Awk is simple and useful enough that it's still worth knowing.
Doesn't every language have regular expressions built in now? Again I still fail to see the point of writing it in Awk when you can write something small and fast in a more powerful and modern language.
It's a higher-level approach than typical scripting languages, and that's why it can be so concise - the model makes a lot of unpacking and looping implicit. It's a DSL for stream-processing problems which are easy phrased as "count these", "transform this into that", etc.
Are you familiar with Prolog? It uses a similar approach, but can match on whole trees (and other complex, nested data structures), not just a list of $N string/numeric tokens. Also, it supports backtracking - at any point, if it reaches a dead end, it can back up arbitrarily and try a different approach. Sometimes slow, but very handy for prototyping.
I agree that using another language than awk makes sense after a few lines, but it's still a sweet spot for 1-5ish line programs. Since awk itself is small enough that a two page cheat sheet is sufficient, it's worth keeping around. Perl (for example) has many nooks and crannies I forget about if I don't use it frequently.
The parent post mentions Prolog, which is a good example, but there are several others worth trying that frequently come up on HN; Scala, Haskell, F#, and Ocaml spring to mind.
I can't speak for Scala, but the PM in Haskell and OCaml is a bit different since it's informed by the static typing. When patterns have variant types (i.e., x is either Foo, Bar, or Baz * int), it also checks for complete coverage. Same general concept, different flavor. Also very useful.
I mentioned Prolog in particular because its emphasis on unification and backtracking make it the most pattern-matching-centric programming language I've seen. Where other languages have pattern matching, it almost is pattern matching.
Also, there are well-known ways to compile pattern specifications into efficient decision trees, so while it's a very expressive abstraction, it's not necessarily an expensive one. If they're being constructed at runtime (as they are in my Lua library), you can generally get a big improvement by just indexing on the patterns' first fields and doing linear search thereafter.
I can say the same for ruby and python (and perl).
From personal experience, as an awk script/program becomes more important - it will evolve with more requirements and it will start to be clunky. It just isn't practical to stick with it since you'll eventually need the features/libraries that the other languages have. Given the choices we have today, why even start with awk?
On the performance side, you can always just use Lua if that's really important.