This study focuses on performance assessment of model predictive control. An MPC-achievable benchmark for the unconstrained case is proposed based on closed-loop subspace identification. Two performance measures can be constructed to evaluate the potential benefit to update the new identified model. Potential benefit by tuning the parameter can be found from trade-off curves. Effect of constraints imposed on process variables can be evaluated by the installed controller benchmark. The MPC-achievable benchmark for the constrained case can be estimated via closed-loop simulation provided that constraints are known. Simulation of an industrial example was done using the proposed method.