First, simulated annealing is a metaheuristic method, it need operators to carry out the changes in the solution, so there is no "simulated annealing alone", it is always coupled with other methods, the better the mix, the better the solutions generated. Second, except for very small instances, a VRP is hardly "solved" to optimality. Regardless, simulated annealing with very simple operators (such as inter-route change and insertion) does not give such good results.