Molecular mechanics (MM) force field models have been demonstrated to have difficulty reproducing certain potential energy surfaces of π-stacked complexes. Here, we examine the performance of the AMBER and CHARMM models relative to high-quality ab initio data across systematic helical parameter scans and typical B-DNA geometries for π-stacking energies of nucleobase dimers. These force fields perform best for typical B-DNA geometries (mean absolute error < 1 kcal mol(-1)), whereas errors typically approach ∼2 kcal mol(-1) for broader potential scans, with maximum errors > 10 kcal mol(-1) relative to high-quality ab initio reference interaction energies. The adequate performance of MM models near minimum energy structures is accomplished through cancellation of errors in various energy terms, whereas large errors at short intermolecular distances are caused by large MM electrostatics errors due to a lack of explicit terms modeling charge penetration effects.