This was just out in PLoS ONE:
Optimization of Genome Engineering Approaches with the CRISPR/Cas9 System, by Li et al. The final figure shows a direct comparison between 50 and 200 bp homology arms for GFP knock-in via CRISPR-mediated HDR; no big surprise, but clearly 200 bp had greater efficiency to 50 bp, by about an order of magnitude. This is useful info for anyone who is designing "large" knock-in cassettes and wondering if it's important to put long arms on the cassettes; although ~60 bases is commonly used for ssDNAs in HDR (i.e. for peptide tag or loxP insertion), it's important to lengthen them for longer cassettes. I suggest ~1 kb. (my gut feeling as of today).