We have performed a comparative study of three widely used classical many-body potentials for silicon (Stillinger-Weber, Tersoff-2, and Tersoff-3) in connection with finite-temperature simulations of the Si(001) surface. Large-scale constant-pressure Monte Carlo simulations are used to examine the reconstruction of the Si(001) surface, formation of antiphase boundaries, and defect structures such as dimer vacancies. The accuracy of the empirical potentials is compared with first-principles methods and with experimental results when possible. We find that good performance in the static limit (i.e:, at T = 0 K) does not ensure the suitability of the potential for finite-temperature simulations. The Stillinger-Weber potential is found to give the best overall performance.