The problem of estimating the error in numerical flow simulations is very important. One case where this is particularly true is in the assessment of turbulence models, where the numerical errors must be known in order to separate them from model errors. In this paper, the three-dimensional steady turbulent flow in a curved rectangular duct has been examined. The Reynolds number was 224000 based on the channel width. At the inlet the flow pattern was already complicated, with a pair of vortices inside the top-wall boundary layer. The results from this case has already been presented (Bergstrom, 1996) but the numerical accuracy and turbulence modeling has not been studied in detail until now. The measurements for the curved channel case were extensive and well suited for checking the turbulence model. The numerical accuracy concerning iterative and grid-convergence was controlled by a grid- and iterative error, respectively. The grid error gives a way to report the number of grid cells needed for a virtually grid-independent solution. A secondary result of the error estimation is that a better approximation to the exact solution is obtained. A Reynolds stress model was used to model the turbulence. The model was seen to be able to capture the most important flow phenomena in the channel.