Accurate geomagnetic field models are crucial to the study of radiation belt phenomena. We quantitatively examine the accuracy of several external models widely in use via the Office National d'Etudes et de Recherche Aérospatiales-Département Environnement Spatial (ONERA-DESP) libraries. We study 2 years characterized by very different space weather conditions, 1996 and 2003. The year 1996, at solar minimum, exhibited many high-speed streams and a few corotating interaction regions but was generally quiet. In contrast, 2003 included the Halloween storm, one of the most intense geomagnetic storms on record caused by a coronal mass ejection. The performance of each model, as measured by prediction efficiency and skill score, is evaluated as a function of magnetospheric conditions (reflected by the geomagnetic index Kp) and magnetic local time (MLT). Not surprisingly, the newer models tend to perform better and interesting comparisons arise between the performances of the models during different periods of the solar cycle and across different Kp and MLT values. For Kp < 4, most models show similar performance, but for higher values, there are large differences between newer and older model performance. As a function of MLT, noticeable dips in the performance of older models are observed near dawn. These dips are suspected to be effects of field-aligned and partial ring currents that are not fully incorporated into the models, but their exact nature is unknown.