Recent work on inference in regressions with heteroscedastic disturbances leads to the possibility of more efficient estimators. The conditions needed to exploit the possibility are discussed. A sampling experiment indicates that the gains can be achieved even in small samples.