Vertebrate mitochondrial cytochrome c oxidase (COX) possesses 10 nuclear-encoded subunits. Six subunits have paralogs in mammals, but the origins and distribution of isoforms among vertebrates have not been analyzed. We used Bayesian phylogenetic analysis to interpret the origins of each subunit, inferring the roles of gene and genome duplications. The paralogous ancestries of five genes were identical throughout the major vertebrate taxa: no paralogs of COX6c and COX7c, two paralogs of COX4 and COX6a, and three paralogs of COX7a. Two genes had an extra copy in teleosts (COX5a, COX5b), and three genes had additional copies in mammals (COX6b, COX7b, COX8). Focusing on early vertebrates, we examined structural divergence and explored transcriptional profiles across zebrafish tissues. Quantitative transcript profiles revealed dramatic differences in transcript abundance for different subunits. COX7b and COX4 transcripts were typically present at very low levels, whereas COX5a and COX8 were in vast excess in all tissues. For genes with paralogs, two general patterns emerged. For COX5a and COX8, there was ubiquitous expression of one paralog, with the other paralog in lower abundance in all tissues. COX4 and COX6a shared a distinct expression pattern, with one paralog dominant in brain and gills and the other in muscles. The isoform profiles in combination with phylogenetic analyses show that vertebrate COX isoform patterns are consistent with the hypothesis that early whole genome duplications in basal vertebrates governed the isoform repertoire in modern fish and tetrapods, though more recent lineage-specific gene/genome duplications also play a role in select subunits.