Prediction of long-term static fatigue for optical fibers under stress requires a model to relate short-term accelerated test results to long-term behavior. The dependence of crack growth on stress intensity is the most fundamental model for this reliability prediction process. Statistical uncertainty for fatigue testing is shown to be significant, but typically smaller than the model uncertainty, which has been neglected in the literature. More research is needed to determine the most appropriate model. It is shown that the differences in allowed stress predictions between models become quite large at long times, especially for fiber of the same strength as that used in the fatigue test. Data-independent conversions of allowed stress from the common power model to other models provide an assessment of the difference between models for various situations. In many applications, the differences are in the range of typical safety factors. However, since model differences are quite large in other applications, universal use of the optimistic power law model is not appropriate, given the limited understanding of fatigue in optical fibers.