A popular hypothesis for human running is that gait mechanics and muscular activity are optimized in order to minimize the cost of transport (CoT). Humans running at any particular speed appear to naturally select a stride length that maintains a low CoT when compared with other possible stride lengths. However, it is unknown if the nervous system prioritizes the CoT itself for minimization, or if some other quantity is minimized and a low CoT is a consequential effect. To address this question, we generated predictive computer simulations of running using an anatomically inspired musculoskeletal model and compared the results with data collected from human runners. Three simulations were generated by minimizing the CoT, the total muscle activation or the total muscle stress, respectively. While all the simulations qualitatively resembled real human running, minimizing activation predicted the most realistic joint angles and timing of muscular activity. While minimizing the CoT naturally predicted the lowest CoT, minimizing activation predicted a more realistic CoT in comparison with the experimental mean. The results suggest a potential control strategy centred on muscle activation for economical running.