The accuracy of six turbulent flow modeling techniques in an unsteady solution is evaluated against experimental data for a square prism in cross flow. The selected models, shear stress transport (SST), SST-SAS, Reynolds stress model (RSM), partially averaged Navier–Stokes (PANS)-SST, detached eddy simulation (DES), and large eddy simulations (LES), models are the same as those presented in Part 1 of this study, which focused on flow in a staggered tube bank. For this geometry, the SST model proved to be effective at capturing the averaged Nusselt values per side of the square with relatively low computational costs. The SST model, however, showed poorer fidelity to the local Nusselt number profile compared to the experimental data. The LES approach provided a more accurate representation of the local Nusselt number but the computational cost was significantly higher. The PANS modification to the SST model did provide a noticeable improvement in accuracy at a reasonable cost while the SAS modification did not see the same improvement. These conclusions are generally consistent with those found for the staggered tube bank in Part 1 of this study. This study can be used as a guide for the industrial user to select a turbulence model for a similar problem with a low Reynolds number and significant flow separation.