Genetic and environmental variation of functional traits within populations might be maintained by natural selection when resource allocation costs (RACs) balance trait benefits. Despite the intuitive appeal of optimization models, empirical tests have failed to support the importance of RACs for plant traits that confer resistance against pests. To address this discrepancy, we modified an early model by allowing the cost function to vary across a resource gradient as predicted for RACs and by assuming that the benefits depend on variation in the pest population for susceptibility. Instead of the intermediate defense optimum of the original model, defenses were predicted to be either high or absent, depending on resource availability and history. This result is not supported by empirical tests for ecological or evolutionary outcomes, including our own examination of glucosinolate toxins from sib-families of Boechera stricta (Brassicaceae) grown across an NPK fertilizer gradient. Although we detected an apparent cost of defense in the absence of herbivores, the cost did not increase as resources became more limiting. Also defense production did not vary across the resource gradient as predicted by the modified model. Thus, a model based on explicit expectations of RACs produced predictions that are not supported. Instead, other kinds of costs, such as ecological (indirect) costs may be more important. Alternatively, general conflicts in gene expression and antagonistic crosstalk among signaling pathways may underlie apparent costs.