Maximum-likelihood detection in MIMO communications amounts to solving a least-squares problem with a constellation (alphabet) constraint. One popular method that can be used to solve this problem is sphere decoding. We show in this letter that by employing a simple stopping criterion, it is possible to significantly reduce the complexity of sphere decoding over a wide range of SNRs, without a noticeable performance degradation. Specifically, simulation results demonstrate that a 10%-90% reduction of the average complexity could be achieved.