This paper proposes a loss-minimizing torque control scheme for induction machines. It is found that during torque transients, the steady-state loss minimization schemes lead to excessively high power losses, as high current peaks are involved. To avoid this, torque transients have to be accounted for in the loss minimization. The flux is controlled such that a desired asymptotic torque response is obtained while the ohmic and iron losses are minimized, thus optimizing efficiency under a constrained output. The dynamic optimization problem is analyzed and found to be very difficult to solve. However, it is shown that with a good approximation, it can be assumed that the optimal flux is an exponential function, and that its time constant can be determined numerically. This heuristic approximation of the dynamic optimum leads to good efficiency improvements. The method is integrated in a predictive control scheme where the optimization is performed online at every sampling step. Experimental results point out the differences to rated operation and to steady-state loss minimization. A flywheel setup demonstrates the additional energy-saving potential in common servo and traction drive applications.