Firstly, we reviewed two extensions of the Erlang multi-rate loss model, whereby we can assess the call-level QoS of telecom networks supporting elastic traffic: (i) the extended Erlang multi-rate loss model, where random arriving calls of certain bandwidth requirements at call setup can tolerate bandwidth compression while in service; and (ii) the connection-dependent threshold model, where arriving calls may have several contingency bandwidth requirements, whereas in-service calls cannot tolerate bandwidth compression. Secondly, we proposed a new model, the extended connection-dependent threshold model. Calls may have alternative bandwidth requirements at call setup and can tolerate bandwidth compression while in service. We proposed a recurrent formula for the efficient calculation of link occupancy distribution and consequently call blocking probabilities, link utilization, and throughput per service class. Furthermore, in the proposed model, we incorporated the bandwidth reservation policy, whereby we can (i) equalize the call blocking probabilities of different service classes, (ii) guarantee specific QoS per service class, and (iii) implement different maximum bandwidth compression/expansion rate per service class so that the network supports both elastic and stream traffic. The accuracy of the new model is verified by simulation. Moreover, the proposed model performs better than the existing models. Finally, we generalize the proposed model by incorporating service classes with either random or quasi-random arrivals.applications whose bandwidth requirement can be compressed, but their service time is not altered (e.g., adaptive audio or video). Calls of fixed bandwidth and service time requirements compose the so-called stream traffic. Despite the fundamental difference between elastic/adaptive traffic and stream traffic, the springboard of call-level modelling of elastic/adaptive traffic is the Erlang multirate loss model (EMLM) [13,14], which is a stream traffic model. In the EMLM, Poisson arriving calls of stream service classes are accommodated to a single link of certain capacity. The available link bandwidth is shared according to the complete sharing (CS) policy [15,16] (calls are accepted for service only when the required bandwidth is available in the link). The calls' service time can be arbitrarily distributed [13]. Stream calls have fixed bandwidth requirement-that is, the assigned bandwidth can change neither during call setup nor during service time. As far as the equilibrium state probabilities are concerned, the EMLM has a product form solution (PFS) ‡ [13]. This fact leads to an accurate calculation of link occupancy distribution and call blocking probabilities (CBP), the key call-level performance metric. Moreover, these calculations are recurrent (Kaufman-Roberts recursion), a feature that broadens the EMLM's applicability range to wired, wireless, or optical networks (e.g., [18][19][20][21][22][23]). The call admission policy has a strong impact on teletraffic models and on the efficie...