The very long baseline interferometry technique allows us to determine the positions of thousands of radio sources using the absolute astrometry approach. I have investigated the impacts of a selection of observing frequencies in a range from 2 to 43 GHz in single-band, dual-band, and quad-band observing modes on astrometric results. I processed seven data sets in a range of 72,000 to 6.9 million observations, estimated source positions, and compared them. I found that source positions derived from dual-band, quad-band, and 23.6 GHz single-band data agree at a level below 0.2 mas. Comparison of independent data sets allowed me to assess the error levels of individual catalogs: 0.05–0.07 mas per position component. Further comparison showed that individual catalogs have systematic errors at the same level. The positions from 23.6 GHz single-band data show systematic errors related to the residual ionosphere contribution. Analysis of source position differences revealed systematic errors along jet directions at a level of 0.09 mas. Network-related systematic errors affect all the data, regardless of frequency. Comparison of position estimates allowed me to derive the stochastic error model that closes the error budget. Based on the collected evidence, I have made a conclusion that the development of frequency-dependent reference frames of the entire sky is not warranted. In most cases dual-band, quad-band, and single-band data at a frequency of 22 GHz and higher can be used interchangeably, which allows us to exploit the strength of a specific frequency setup for given objects. Mixing observations at different frequencies causes errors not exceeding 0.07 mas.