Genetic programming (GP) based symbolic regression is a stochastic, high-variance algorithm. Its sensitivity to changes in training data is a drawback for practical applications. In this work, we analyze empirically the variance of GP models on the PennML benchmarks. We measure the spread of model predictions when models are trained on slightly perturbed data. We compare the spread of models from two GP variants as well as linear, polynomial and random forest regression models. The results show that the spread of models from GP with local optimization is significantly higher than that of all other algorithms. As a side effect of our analysis, we provide evidence that the PennML benchmark contains two groups of instances (Friedman and real-world problem instances) for which GP performs significantly different.