We present the results of the fully automated CAFASP3 experiment, which was carried out in parallel with CASP5, using the same set of prediction targets. CAFASP participation is restricted to fully automatic structure prediction servers. The servers' performance is evaluated by using previously announced, objective, reproducible and fully automated evaluation methods. More than 60 servers participated in CAFASP3, covering all categories of structure prediction. As in the previous CAFASP2 experiment, it was possible to identify a group of 5-10 top performing independent servers. This group of top performing independent servers produced relatively accurate models for all the 32 "Homology Modeling" targets, and for up to 43% of the 30 "Fold Recognition" targets. One of the most important results of CAFASP3 was the realization of the value of all the independent servers as a group, as evidenced by the superior performance of "meta-predictors" (defined here as predictors that make use of the output of other CAFASP servers). The performance of the best automated meta-predictors was roughly 30% higher than that of the best independent server. More significantly, the performance of the best automated meta-predictors was comparable with that of the best 5-10 human CASP predictors. This result shows that significant progress has been achieved in automatic structure prediction and has important implications to the prospects of automated structure modeling in the context of structural genomics.