Computationally-expensive regional climate models (RCM) are increasingly being used to generate local climate data for climate change impact studies. These studies usually process RCM output to remove errors in the simulated climate. However, this paper investigates the suitability of raw output from simulations of a single RCM as input to a biophysical impact model. Our study analyses errors in wheat yields simulated for New South Wales (NSW), Australia by the Agricultural Production Systems Simulator (APSIM) model forced with output from 2 RCM simulations with horizontal resolutions of approximately 50 and 10 km over NSW and with output from the global climate model (GCM) simulation that they downscale. Overall, across the NSW wheat belt, the ~50 km simulation has a better simulation of mean yields for the 1990-2010 period than the GCM simulation, and the ~10 km simulation has a better simulation than the ~50 km simulation. The average mean yield from APSIM simulations forced with observations is 3.5 t ha-1. The average magnitudes of errors in mean yields for the GCM, ~50 km and ~10 km simulations are 1.2, 1.0 and 0.5 t ha-1 respectively. We suggest that the improvement in the simulation of mean yields with increasing climate model resolution is largely due to an improvement in the simulation of mean rainfall totals for the growing season. However, for a given value of mean growing season total rainfall, all 3 climate model simulations have a climate that is more conducive to high yields than the observed climate. This difference must be due to errors in other aspects of the simulated climates.