The present paper deals with loss minimization for speed controlled squirrel-cage induction machines. In particular, a flatness-based online trajectory planning algorithm for the rotor flux is introduced that ensures loss optimization in transient operating regimes. Therefore, a matched model of the losses being polynomial in the flux is used. For a trajectory of the torque which is a polynomial function of time it is possible to obtain the parameters for a polynomial prototype of the flux trajectory such that a loss-optimal regime can be approximately reached. The generation of the trajectories is shown exemplary for linear speed transitions. A flux-oriented cascaded control scheme with a flatness-based feed-forward ensures tracking of the nominal trajectories. Experiments with different linear speed transitions show significant improvements of the efficiency in comparison to the operation with rated flux and a flux optimized for steady state.