no simple answer (there may be different good clusterings competing at different granularity levels, for instance and the "optimality" depends on what you plan todo with your clusters !) but a rather efficient method relying on a recursive adaptation of permutation tests is described in
Thank you sir, I am trying to cluster morphological variant words. tokens are in thousands so by just visualizing dendogram we can not decide from where we need to cut down the dendo gram to get proper clusters.