One obstacle in the translation of advances in cancer research into the clinic is a deficiency of adequate preclinical models that recapitulate human disease. Patient-derived xenograft (PDX) models are established by engrafting patient tumor tissue into mice and are advantageous because they capture tumor heterogeneity. One concern with these models is that selective pressure could lead to mutational drift and thus be an inaccurate reflection of patient tumors. Therefore, we evaluated if mutational frequency in PDX models is reflective of patient populations and if crucial mutations are stable across passages. We examined KRAS and PIK3CA gene mutations from pancreatic ductal adenocarcinoma (PDAC) (n = 30) and colorectal cancer (CRC) (n = 37) PDXs for as many as eight passages. DNA was isolated from tumors and target sequences were amplified by polymerase chain reaction. KRAS codons 12/13 and PIK3CA codons 542/545/1047 were examined using pyrosequencing. Twenty-three of 30 (77%) PDAC PDXs had KRAS mutations and one of 30 (3%) had PIK3CA mutations. Fifteen of 37 (41%) CRC PDXs had KRAS mutations and three of 37 (8%) had PIK3CA mutations. Mutations were 100 per cent preserved across passages. We found that the frequency of KRAS (77%) and PIK3CA (3%) mutations in PDAC PDX was similar to frequencies in patient tumors (71 to 100% KRAS, 0 to 11% PIK3CA). Similarly, KRAS (41%) and PIK3CA (8%) mutations in CRC PDX closely paralleled patient tumors (35 to 51% KRAS, 12 to 21% PIK3CA). The accurate mirroring and stability of genetic changes in PDX models compared with patient tumors suggest that these models are good preclinical surrogates for patient tumors.