We discuss the accuracy of the time-discrete phase noise model described by a random walk with symbol-period spaced Gaussian increments. While this model is widely used for its simplicity, it is strictly valid in the slow phase regime only. It is customary to consider this model as a worst case, but a clear understanding of the phase dynamics which can be reliably represented may not be readily available. We address this problem by comparing the symbol-period spaced model with fractional-period spaced ones, which more reliably describe the physical system, in terms of the achievable information rate of the resulting phase noise channels. We show that indeed the symbol-period spaced model is a worst case that may offer a good modeling accuracy under a wide range of phase dynamics.