Application of continuum electrostatic calculations to molecular modeling requires an accurate description of molecular charge distributions, typically as partial atomic charges. While for some systems, such as biological macromolecules, sets of charges have been parametrized on the basis of experimental data, for many other cases, ab initio methods of charge determination may be preferred. Presented here is a comprehensive evaluation of the performance of numerous methods for the ab initio determination of partial atomic charges in continuum electrostatic calculations. Charges were computed using several methods based both on fitting electrostatic potentials and on population analysis, and using various levels of theory ranging from semiempirical quantum mechanical methods through relatively high level ab initio quantum mechanical methods. All charge distributions were evaluated in terms of their ability to reproduce experimental free energies of solvation in the context of a continuum solvation model. Two sets of test molecules were used, one derived from the groups seen in proteins, and the other a more diverse set of neutral organic molecules. The results indicate that there are clearly preferred methods for determining charges and, conversely, that there are highly unsuitable methods. The agreement with experiment does not increase monotonically with increasingly accurate levels of theory, although the lowest level methods do perform particularly poorly. None of the methods performed uniformly well across all molecule types; the top performing methods tended to give charge magnitudes in the middle of the observed range, but both the under- and overpolarized charge distributions perform better for certain systems. The frequently used HF/6-31G* level of quantum mechanics did very well, ranking among the top methods, particularly when coupled with the Merz-Singh-Kollman charge fitting scheme or a restrained fit based on this scheme. For methods at a relatively high level of theory, the charges derived from B3LYP/6-311G*(+) potentials performed the best, while the computationally inexpensive B3LYP/4-31G derived charges provided the best performance for a fairly low level method.