Genomic selection (GS) is a powerful tool for improving genetic gain in maize breeding. However, its routine application in large-scale breeding pipelines is limited by the high cost of genotyping platforms. Although sequencing-based and array-based genotyping platforms have been used for GS, few studies have compared prediction performance among platforms. In this study, we evaluated the predictabilities of four agronomic traits in 305 maize hybrids derived from 149 parental lines subjected to genotyping by sequencing (GBS), a 40K SNP array, and target sequence capture (TSC) using eight GS models. The GBS marker dataset yielded the highest predictabilities for all traits, followed by TSC and SNP array datasets. We investigated the effect of marker density and statistical models on predictability among genotyping platforms and found that 1K SNPs were sufficient to achieve comparable predictabilities to 10K and all SNPs, and BayesB, GBLUP, and RKHS performed well, while XGBoost performed poorly in most cases. We also selected significant SNP subsets using genome-wide association study (GWAS) analyses in three panels to predict hybrid performance. GWAS facilitated selecting effective SNP subsets for GS and thus reduced genotyping cost, but depended heavily on the GWAS panel. We conclude that there is still room for optimization of the existing SNP array, and using genotyping by target sequencing (GBTS) techniques to integrate a few functional markers identified by GWAS into the 1K SNP array holds great promise of being an effective strategy for developing desirable GS breeding arrays.