Optimal input design is considered for discriminating effectively between two rival autoregressive models when the amplitude of the system output has to be regulated within a certain tolerance limit with high certainty. This constraint is more appropriate than a power constraint when an extremely large system output may cause hazardous conditions in the system. First, conditions for the optimal input are derived based on the Ds-criterion, which corresponds to the power of the likelihood ratio test. Then two approaches are presented to construct the optimal input satisfying the conditions: one is based on the idea of a Chebyshev system, and the other is an autoregressive recursion approach. Numerical simulations illustrate the applicability of the proposed optimal input for autoregressive model discrimination.