The study compares yield, biomass and water use (WU) for maize, sorghum and millet simulated using three crop models of varying complexity: AquaCrop, DSSAT and the SIMPLE model. A standard set of crop parameters was used to develop crop files for all three models. Similar soil, climate and management descriptions from the Ukulinga Research Farm were used across the models. The performance of the three models was observed to be statistically different. Based on the mean bias error, all models overestimated yield, but the lowest overestimation was with AquaCrop (0.22 t/ha), followed by DSSAT (0.24 t/ha) and the SIMPLE model (0.69 t/ha). Other statistical indicators, namely, RMSE and R2, illustrated that the simulation of yield and WP in AquaCrop was more satisfactory than DSSAT and the SIMPLE model. The study confirms that DSSAT requires relatively more input data but does not always perform more satisfactorily. Before their application, it is essential to calibrate crop growth parameters for local conditions or use parameters from local field studies when applying complex crop models such as DSSAT specifically for marginal environments, such as South Africa. On the other hand, AquaCrop performed reasonably well with minimal input requirements, confirming its application in data-limited and marginal environments. However, it is recommended that there must be calibration for all the models using inputs specific to locations.