For this CRISPR question I am going back to this paper: Yang et al, Optimization of scarless human stem cellgenome editing, Nucleic Acids Research, 2013, 1–13 (from the Church lab). Although this paper had a lot of TALEN data it had comparative data for CRISPR-mediated editing. In this case, a 2-bp mismatch was engineered into the CCR5 locus in human iPS cells.
First of all, from now on I will use the term "ssODN" to refer to single-stranded donor oligonucleotides. (Hooray, more jargon!) In Fig. 3d of Yang et al, they presented a nice series of data in which they varied both the length and the strandedness of the ssODN used for editing in conjunction with a single gRNA, which was held constant of course. The mismatch was within the CRISPR target and was always positioned in the middle of the ssODN. However, ssODN length varied from 50 to 110 nt. (Strangely, the top panel has longer ssODNs also drawn schematically but no data for those was shown). In addition, ssODNs corresponding to either the same strand or the complementary strand to the gRNA were tested.
Now, if you're like me, you might naively assume that the complementary-stranded ssODN would be worse in mediating editing because it might base-pair with the gRNA itself, preventing the gRNA from functioning properly. Sounds reasonable? Turns out, the opposite was true - at least for this individual target. Maximal efficiency of editing was achieved with the complementary ssODN at about 70 nt length, with an absolute efficiency of ~1.5% - not too shabby considering it's iPS cells and without any selection. Interestingly, when non-complementary ssODN (i.e., same strand as the gRNA) were used, the efficiency never reached that efficiency, but it did increase with length up to the maximum tested length (110 nt) where it reached about 0.5%. At this length, it was basically the same whether the complementary or non-complementary ssODN was used.
What should one take from this? Well, the trouble with these sorts of tests is that when you test certain parameters you have to keep the other parameters fixed. In this case the gRNA was kept constant. So the peak in efficiency at 70 nt with the complementary ssODN might be a peculiarity of that particular sequence - perhaps it forms a secondary structure that just happens to inhibit the competing NHEJ pathway, for example?
Also, there is a decent theme forming that for mouse oocyte injections, ssODNs should have homology arms of about 60 nt. This puts the minimum length at 120 nt.
P.S. ...another thought added 12/15/14: Note that for the above observation that max. editing efficiency was achieved with a complementary-strand ssODN of ~70 nt, this means the homology arms were each only ~35 nt long. In fact, a 50 nt ssODN was as efficient as a 90 nt ssODN, meaning ~25 nt homology arms were also workable with a complementary ssODN. But the non-complementary ssODN worked better with longer arms. See Fig. 3 from the Yang paper.