A recent question on ab initio that I answered involved touching on the limitations of ab initio modelling. A review from 2009 put forward that typically only in the order of 100-110 residues could be reliably modelled.
7 years after that paper: I just attended a pre-publication presentation in which a 400+ residue structure had been modelled using I-TASSER. The speaker claimed it was reliable and it did fit well in the biological context, but it still needed experimental validation. Unfortunately they had no in silico validation scores to hand so I had no handle on how seriously I should take the model.
Ultimately, my question is how many residues can I-TASSER (other modern ab initio brands available!) reliably handle for a typical globular domain with today's technology?