DiffServ network model, which is a scalable framework for providing Quality of Service to applications, defines multiple service classes each with different level of priority in queuing of network routers. Recent study shows that the global treatment of premium class as the most prioritized service class strongly affects the performance of low priority classes. This negative impact becomes more acute in the core of the network where the traffic is aggregated. In order to alleviate the inter-class effect, the routing algorithm used to compute the premium paths must minimize the load of high priority class over the network links. In this paper, we propose a new routing architecture which allows the core routers to use multipath scheme for forwarding of arriving premium packets. Simulation results show that our approach outperforms several earlier algorithms in terms of maximum premium congestion on the network.