This paper introduces a modification to the affine GARCH model of Heston and Nandi (2000). The new model is designed to allow for a non-zero lower bound for the variance achieved by adding two parameters to the existing model. The affine structure of the moment generating function is preserved at the level of variance, while an approximation is studied for log prices. The construction resembles the shifted continuous-time Heston (1993) model. Maximum likelihood estimation is performed on real data, and the model is shown to improve on the fitting of the implied volatility surface particularly for deep out-of-the-money options.