This talk reviews the predictability and prediction skill of CMIP5 models on multi-year time scales. New statistical optimization techniques substantially clarify the space-time structure of predictability and reveal that virtually all of the predictability of annual means can be explained by the following components: the response to natural and anthropogenic forcing, the Atlantic Multidecadal Oscillation (AMO), the Pacific Decadal Oscillation (PDO), the Southern Annular Mode (SAM), and ENSO. The technique also reveals predictability of seasonal mean temperature and precipitation over land on multi-year time scales, apparently for the first time. An empirical forecast system based on these predictable components is verified to have skill on multi-year time scales over the entire twentieth century observational record. The predictability of the AMO varies widely among CMIP5 models and is difficult to validate based on 150 years of observations. The relation between AMO and other variables in the atmosphere and ocean are examined to understand the cause for widely varying predictability estimates.