The Cython example in that link is actually not a fair comparison, since it still forces the numpy ndarray type in the signature.
Instead it should use typed memoryviews [0], which are faster and can avoid more cases that will rely on the GIL accidentally (such as when an ndarray has to be treated as a Python object).
but you should link to some docs / examples / tutorials in this thread because the release notes doesn't do justice of what Pythran can do.
[1] https://github.com/SimonDanisch/julia-challenge/pull/4