Prediction of the phenotypes for a set of genotypes across multiple environments is a fundamental task in any plant breeding program. Genomic prediction (GP) can assist selection decisions by combining incomplete phenotypic information over multiple environments (MEs) with dense sets of markers. We compared a range of ME-GP models differing in the way environment-specific genetic effects were modeled. Information among environments was shared either implicitly via the response variable, or by the introduction of explicit environmental covariables. We discuss the models not only in the light of their accuracy, but also in their ability to predict the different parts of the incomplete genotype × environment interaction (G × E) table: (Gt; Et), (Gu; Et), (Gt; Eu), and (Gu; Eu), where G is genotype, E is environment, both tested (t; in one or more instances) and untested (u). Using the ‘Steptoe’ × ‘Morex’ barley (Hordeum vulgare L.) population as an example, we show the advantage of ME-GP models that account for G × E. In addition, for our example data set, we show that for prediction in the most challenging scenario of untested environments (Eu), the use of explicit environmental information is preferable over the simpler approach of predicting from a main effects model. Besides producing the most general ME-GP model, the use of environmental covariables naturally links with ecophysiological and crop-growth models (CGMs) for G × E. We conclude with a list of future research topics in ME-GP, where we see CGMs playing a central role.