If your original or logarithmically-transformed variables are I(1) and cointegrated, then VECM is the correct model. Unrestricted VAR in levels is misspecified due to missing restrictions because of cointegration, while VAR for variables in first differences is misspecified because it omits the error correction terms (ECTs).
Now whether VECM is preferred over VAR for a particular objective is not clear cut. For example, if the ECTs are estimated with high variance, their inclusion may result in worse forecasting performance than that of VAR in first differences (in which the ECTs are omitted). This is a special case of the more general fact that correctly specified models with estimated (rather than known) parameters need not be the best models for forecasting, which is because of estimation imprecision.