This paper presents an accurate stochastic model for transfer latency of short-lived Web-like TCP flows with random packet losses. Our model characterizes a data transfer in alternating cycles, with TCP state information carried over from one cycle to the next. Simulation experiments show that our model matches simulation results for short-lived flows better than earlier TCP models, and fits long-lived TCP flows as well. Our model is then extended to estimate transfer times for CATNIP TCP, which is shown to be 5-42% faster than TCP Reno, depending on transfer size and packet loss ratio.