Both the SARS-CoV-2 virus and its mRNA vaccines depend on RNA polymerases (RNAP)1,2; however, these enzymes are inherently error-prone and can introduce variants into the RNA3. To understand SARS-CoV-2 evolution and vaccine efficacy, it is critical to identify the extent and distribution of errors introduced by the RNAPs involved in each process. Current methods lack the sensitivity and specificity to measure de novo RNA variants in low input samples like viral isolates3. Here, we determine the frequency and nature of RNA errors in both SARS-CoV-2 and its vaccine using a targeted Accurate RNA Consensus sequencing method (tARC-seq). We found that the viral RNA-dependent RNAP (RdRp) makes ~1 error every 10,000 nucleotides – higher than previous estimates4. We also observed that RNA variants are not randomly distributed across the genome but are associated with certain genomic features and genes, such as S (Spike). tARC-seq captured a number of large insertions, deletions and complex mutations that can be modeled through non-programmed RdRp template switching. This template switching feature of RdRp explains many key genetic changes observed during the evolution of different lineages worldwide, including Omicron. Further sequencing of the Pfizer-BioNTech COVID-19 vaccine revealed an RNA variant frequency of ~1 in 5,000, meaning most of the vaccine transcripts produced in vitro by T7 phage RNAP harbor a variant. These results demonstrate the extraordinary genetic diversity of viral populations and the heterogeneous nature of an mRNA vaccine fueled by RNAP inaccuracy. Along with functional studies and pandemic data, tARC-seq variant spectra can inform models to predict how SARS-CoV-2 may evolve. Finally, our results may help improve future vaccine development and study design as mRNA therapies continue to gain traction.