Not really so much. You just target an area where C is not very good and you can do it. For bulk threaded stuff using lots of I/O, Java can be "faster than C." For extremely distributed code, Erlang can be "faster than C."

These claims leverage the fact that the fastest implementation in C is incredibly difficult to write. Usually we discard that difficulty when comparing languages, but as time goes on the costs of developing specific distributed, multi-threaded, and/or reliable systems in pure C becomes so outrageous that it stops making sense to ignore, so we start comparing against less correct implementations.

It's not that much of a claim, if they implicit parallelization works well. Also some functional languages are getting very nice optimizations like deforrestions / fusions that make them really fast.

And it doesn't actually do any code generation yet.

Good point. Also how can you measure this speed when talking about C? First of all there's not only one single implementation of the compiler, and second thing for available architecture you may have different performances basing on the quality of the backend...

I find that kind of claim pretty annoying.