Over the last year it has become apparent that advanced machine learning models are capable to compete with and even outperform conventional numerical weather prediction models. One of these models is GraphCast, developed by Google, and able to produce deterministic forecasts of hundreds of weather variables under one minute at state-of-the-art accuracy. These skills were learned by GraphCast during an extensive training at the heart of which lies the minimization of a loss function. Given this key role, understanding the model’s sensitivity to the loss function is crucial. In this paper we present a comparative analysis of GraphCast’s performance when trained on different loss functions, where we retrain GraphCast employing the mean absolute error and the log-cosh function next to the benchmark mean squared error. We assess the overall impact of different loss functions by calculating various error metrics and demonstrate the influence this choice has regarding the accuracy of weather forecasts across various regions and lead times.