Corrosion penetration of steel depends on the chemical composition of the material, the environment at the exposure site, and the extent to which debris accumulates on the surface of the specimen. Corrosion penetration data have traditionally been modeled with a power function over the full length of exposure time. The power model lacks the flexibility to describe a variety of behaviors and often underpredicts corrosion penetration at the end of the service life of a structure. Composite models are shown herein to better represent corrosion penetration data. Thirty-eight sets were fitted with power and composite models. In all cases, the composite model fitted the data at least as well as the power model, and in most cases better judging by goodness-of-fit statistics and errors between the predicted and measured penetrations. Errors in penetration predicted with the power model often show large local biases, suggesting the model structure is incorrect. Local biases also serve as a warning that projections of penetration to the end of the service life may be highly inaccurate. Conversely, composite models have unbiased coefficients and better represent the factors that govern corrosion penetration, thus giving greater confidence in service life projections.