After running some tests it does look like they carefully cherry picked the publicised performance, what I saw was a nearly random pattern of relative compression compared to lzma and deflate. In one test deflate beat it by >10%!
The lzma comparison was more interesting, with the quality at usable settings it was comfortably faster with roughly the same compression ratio. That's pretty surprising given the use of huffman coding. Will definitely be trialling it some more, lzma just isn't fast enough for us (neither is lzham), this might just be.
A future version with arithmetic coding could be very interesting, once the current patent woes on ANS are dealt with.