A few examples:
http://www.sigir.org/forum/F2001/sigirFall01Letters.html
http://www.math.columbia.edu/~woit/wordpress/?p=442
http://www.math.columbia.edu/~woit/wordpress/?p=581
JMLR, at least, has gone on to successfully eclipse the journal, Machine Learning, that it was intended to replace (about 3x the impact factor).
I do notice that the examples I find are all in computer science and mathematics, and the new journals have basically zero budgets (and like it that way) and don't charge authors any fees. Is this because in CS/math a common expectation is that the author can use LaTeX, produce their own figures, and submit a print-ready PDF, whereas in other fields authors expect significant formatting work to be done by the journal?