We propose stochastic packet level models for the NewReno and Tahoe versions of TCP during persistent data transmission. It is shown how throughput and the risk of experiencing time-out depend on the packet error probability. We show how this error rate could be interpreted under different variants of Active Queue Management like RED and DropTail. The models clearly separate aspects of the flow control dynamics that are fundamental for generic TCP under general networking topologies from those that are highly dependent on the particular choice of AQM. Modeling the AQM dependent correlation structure in packet losses is shown to have a big impact on TCP throughput and time-out probability implying the importance of more measurement based studies on the structure of packet loss correlation in the real Internet. We use a previously introduced technique to approximate window sizes and cycle lengths by continuous quantities and extend earlier derived renewal-reward arguments to be valid for NewReno as well as for the originally modeled Tahoe. Important aspects of the flow control such as fast retransmit, fast recovery and time-outs are considered. The ns-2 simulator is used to validate and verify the proposed models.