At a genomic scale, the patterns that have shaped molecular evolution are believed to be largely heterogeneous. Consequently, comparative analyses should use appropriate probabilistic substitution models that capture the main features under which different genomic regions have evolved. While efforts have concentrated in the development and understanding of model selection techniques, no descriptions of overall relative substitution model fit at the genome level have been reported. Here, we provide a characterization of best-fit substitution models across three genomic data sets including coding regions from mammals, vertebrates, and Drosophila (24,000 alignments). According to the Akaike Information Criterion (AIC), 82 of 88 models considered were selected as best-fit models at least in one occasion, although with very different frequencies. Most parameter estimates also varied broadly among genes. Patterns found for vertebrates and Drosophila were quite similar and often more complex than those found in mammals. Phylogenetic trees derived from models in the 95% confidence interval set showed much less variance and were significantly closer to the tree estimated under the best-fit model than trees derived from models outside this interval. Although alternative criteria selected simpler models than the AIC, they suggested similar patterns. All together our results show that at a genomic scale, different gene alignments for the same set of taxa are best explained by a large variety of different substitution models and that model choice has implications on different parameter estimates including the inferred phylogenetic trees. After taking into account the differences related to sample size, our results suggest a noticeable diversity in the underlying evolutionary process. All together, we conclude that the use of model selection techniques is important to obtain consistent phylogenetic estimates from real data at a genomic scale.