To delay evolution of pest resistance to transgenic crops producing insecticidal proteins from Bacillus thuringiensis (Bt), the "pyramid" strategy uses plants that produce two or more toxins that kill the same pest. In the United States, this strategy has been adopted widely, with two-toxin Bt cotton replacing one-toxin Bt cotton. Although two-toxin plants are likely to be more durable than one-toxin plants, the extent of this advantage depends on several conditions. One key assumption favoring success of two-toxin plants is that they kill insects selected for resistance to one toxin, which is called "redundant killing." Here we tested this assumption for a major pest, Helicoverpa zea, on transgenic cotton producing Bt toxins Cry1Ac and Cry2Ab. Selection with Cry1Ac increased survival on two-toxin cotton, which contradicts the assumption. The concentration of Cry1Ac and Cry2Ab declined during the growing season, which would tend to exacerbate this problem. Furthermore, analysis of results from 21 selection experiments with eight species of lepidopteran pests indicates that some cross-resistance typically occurs between Cry1A and Cry2A toxins. Incorporation of empirical data into simulation models shows that the observed deviations from ideal conditions could greatly reduce the benefits of the pyramid strategy for pests like H. zea, which have inherently low susceptibility to Bt toxins and have been exposed extensively to one of the toxins in the pyramid before two-toxin plants are adopted. For such pests, the pyramid strategy could be improved by incorporating empirical data on deviations from ideal assumptions about redundant killing and cross-resistance.