Galactic sized gravitational lenses are simulated by combining a cosmological N-body simulation and models for the baryonic component of the galaxy. The lens caustics, critical curves, image locations and magnification ratios are calculated by ray-shooting on an adaptive grid. When the source is near a cusp in a smooth lens' caustic the sum of the magnifications of the three closest images should be close to zero. It is found that in the observed cases this sum is generally too large to be consistent with the simulations implying that there is not enough substructure in the simulations. This suggests that other factors play an important role. These may include limited numerical resolution, lensing by structure outside the halo, selection bias and the possibility that a randomly selected galaxy halo may be more irregular, for example due to recent mergers, than the isolated halo used in this study. It is also shown that, with the level of substructure computed from the N-body simulations, the image magnifications of the Einstein cross type lenses are very weak functions of source size up to $\sim 1\kpc$. This is also true for the magnification ratios of widely separated images in the fold and cusp caustic lenses. This means that selected magnification ratios for different the emission regions of a lensed quasar should agree with each other, barring microlensing by stars. The source size dependence of the magnification ratio between the closest pair of images is more sensitive to substructure.

28 pages, 2 tables and 14 figures. Accepted to MNRAS