Recently published images of the Earth's mantle are characterized by a nominal resolution much higher than that used in previous studies, where substantially different techniques were employed. The agreement between such high-resolution and low-resolution images often seems very poor. In an attempt to determine the reason for this discrepancy, we analyze how the choice of inversion algorithm (exact or iterative), regularization (norm or roughness minimization), and parameterization (spherical harmonics up to a variable degree, blocks) affects a tomographic model. We also investigate the effects of the varying density of the data coverage on the final solution. In our experiments we employ two seismic data sets: Rayleigh wave phase velocity at 75 s period and P wave travel times. We construct a new model of P velocity in the mantle (BDP98) based on the International Seismological Center bulletins 1964--1992. We use our findings in an evaluation of recent mantle models, including our own, focusing on similarities and discrepancies between models of different nominal resolution. In all the models the long-wavelength component is the most stable. However, consistent high-resolution details, probably corresponding to features of the real Earth, are also seen. In general, we conclude that most of the differences between existing tomographic models derive from the arbitrary choices made in the process of defining and solving the inverse problem, rather than from actual errors or approximations. ¿ 1999 American Geophysical Union