As another data point, in my area, artificial intelligence, the open-access journals have overtaken the closed-access journals and don't charge any publishing fees. They generally do this by, quite literally, having close to zero expenses, and covering the rest from donations/sponsors.
For example, the Journal of Machine Learning Research is now the most prestigious ML journal, and it has a $0 budget: it's hosted online on donated servers from MIT (http://jmlr.csail.mit.edu/), authors are expected to deliver publication-ready PDFs via LaTeX, administrative work is done by volunteer editors and students, and archival copies are printed off as print-on-demand by a third party publisher.
The Journal of Artificial Intelligence Research, the top-ranked journal in general AI, runs a somewhat more institutionalized operation (it actually has a few paid staff), but pays for it out of sponsors rather than author fees: http://www.jair.org/
I confess I'm not familiar enough with academic publishing to know why there's such a large difference in cost structures, and why other areas can't do AI-style free-and-open-access journals.