Abstract Inference of deep phylogenies has almost exclusively used protein rather than DNA sequences based on the perception that protein sequences are less prone to homoplasy and saturation or to issues of compositional heterogeneity than DNA sequences. Here, we analyze a model of codon evolution under an idealized genetic code and demonstrate that those perceptions may be misconceptions. We conduct a simulation study to assess the utility of protein versus DNA sequences for inferring deep phylogenies, with protein-coding data generated under models of heterogeneous substitution processes across sites in the sequence and among lineages on the tree, and then analyzed using nucleotide, amino acid, and codon models. Analysis of DNA sequences under nucleotide-substitution models (possibly with the third codon positions excluded) recovered the correct tree at least as often as analysis of the corresponding protein sequences under modern amino acid models. We also applied the different data-analysis strategies to an empirical dataset to infer the metazoan phylogeny. Our results from both simulated and real data suggest that DNA sequences may be as useful as proteins for inferring deep phylogenies and should not be excluded from such analyses. Analysis of DNA data under nucleotide models has a major computational advantage over protein-data analysis, potentially making it feasible to use advanced models that account for among-site and among-lineage heterogeneity in the nucleotide-substitution process in inference of deep phylogenies.