Climate-change scenarios predict increased scarcity of water available for agriculture in irrigated regions. To design and assess cropping system adaptations to more frequent droughts, soil-crop models are useful and efficient. The variety of models available makes it difficult to select one that is robust under a wide range of agro-pedoclimatic contexts and effective with limited input data and standard values. Here, we compared performances of two soil-crop models of contrasting complexity: STICS, a mechanistic model, and AqYield, which is much simpler and more empirical. For this purpose, predicted soil water contents and yields were compared to independent data acquired for three spring crops (maize, sunflower, and sorghum) at four sites in a summer-water-deficient region in southwestern France. Crops were grown under several irrigation strategies, from rainfed to full irrigation conditions. Both models tended to predict yields satisfactorily, but more accurately for maize, intermediate for sorghum and less accurately for sunflower. They accurately ranked situations according to crop, soil and irrigation strategy, but failed to rank inter-annual variability. Both AqYield and STICS predicted much of the variability observed in soil available water content (AWC) under maize and sorghum. Predictions were less accurate, although satisfactory, for sunflower. STICS underpredicted AWC under sorghum, but was generally more accurate than AqYield in situations with low water stress. AqYield was more accurate for high levels of water stress, but tended to overpredict AWC. Yearly dynamics of AWC were evaluated with a novel expert method. Both models accurately represented these dynamics in more than 60% of cases. Overall, we demonstrated that both models sufficiently predicted yield and water balance; however, STICS is more appropriate when other limiting processes need to be simulated.