$Om(z)$ is a diagnostic approach to distinguish dark energy models. However, there are few articles to discuss what is the distinguishing criterion. In this paper, firstly we smooth the latest observational $H(z)$ data using a model-independent method -- Gaussian processes, and then reconstruct the $Om(z)$ and its fist order derivative $mathcal{L}^{(1)}_m$. Such reconstructions not only could be the distinguishing criteria, but also could be used to estimate the authenticity of models. We choose some popular models to study, such as $Lambda$CDM, generalized Chaplygin gas (GCG) model, Chevallier-Polarski-Linder (CPL) parametrization and Jassal-Bagla-Padmanabhan (JBP) parametrization. We plot the trajectories of $Om(z)$ and $mathcal{L}^{(1)}_m$ with $1 sigma$ confidence level of these models, and compare them to the reconstruction from $H(z)$ data set. The result indicates that the $H(z)$ data does not favor the CPL and JBP models at $1 sigma$ confidence level. Strangely, in high redshift range, the reconstructed $mathcal{L}^{(1)}_m$ has a tendency of deviation from theoretical value, which demonstrates these models are disagreeable with high redshift $H(z)$ data. This result supports the conclusions of Sahni et al. citep{sahni2014model} and Ding et al. citep{ding2015there} that the $Lambda$CDM may not be the best description of our universe.