The performance of the complete active space second-order perturbation theory (CASPT2) to accurately predict magnetic coupling in a wide series of molecules and solid state compounds has been established. It is shown that CASPT2, based on a reference wave function that only includes the effects described by the Anderson or Hay-Thibeault-Hoffmann model, correctly reproduces all experimental trends. For a complete quantitative agreement with experiment (or with accurate results arising from the difference dedicated configuration interaction method), it is necessary to include effects that go beyond the Anderson model in the reference wave function. The CASPT2 method is computationally less demanding than CI based methods and, hence, allows us to extend the study of magnetic coupling parameters to larger molecules or systems with elevated spin moments. Moreover, CASPT2 provides a reliable and accurate alternative to density functional based methods that require the use of a broken symmetry approach.