For low heritability traits, large reference populations are required to achieve high reliability of genomic EBV (GEBV). By including genotyped and non-genotyped animals simultaneously in the evaluation, the single-step GBLUP (ssGBLUP) has the potential to yield more accurate and less biased evaluations. The aim of this study was to compare the reliability and bias of genomic predictions for various workability and reproductive traits in Holstein cattle using 2-step GBLUP (tsGBLUP) and ssGBLUP. A total of 33,568 bulls and 6,849 cows were genotyped. Genomic predictions were assessed using genotypes only for bulls or for bulls and cows under different approaches: (1) GEBV estimated by ssGBLUP with a default blending of 5% of the pedigree relationship matrix among genotyped animals (A₂₂) and 95% of the genomic relationship matrix (G); (2) Direct genomic value (DGV) estimated by tsGBLUP where 5% or 20% of A₂₂ was blended into G; (3) GEBVi as an index combining EBV and DGV, where the latter was obtained from tsGBLUP with 5% or 20% blending. Regular BLUP without genomic information were also carried out and EBV served as benchmark for comparisons. Reliabilities were obtained with forward prediction, following the Interbull validation method. Validation bulls had at least 50 daughters in 2014. Including genomic information improved reliability, on average, by 14.5 pts for ssGBLUP and 12 pts for tsGBLUP compared with BLUP. Overall, ssGBLUP predictions had 3.3 pts greater reliabilities and were 0.15 pts less biased compared with tsGBLUP. When a 20% blending was used in stGBLUP, predictions were less biased, but no differences in reliability were observed compared with a 5% blending. Adding genotypes for cows had a small, positive impact of 1.5 pts in reliability for ssGBLUP and 1.6 pts for tsGBLUP; bias was reduced by 0.002 and 0.023, respectively. When genomic information is available for cows, predictions for both workability and reproductive traits can be slightly improved. Single-step GBLUP leads to more accurate and less biased predictions compared with 2-step GBLUP.