© 2019 Elsevier Inc.A significant cost of traffic congestion is unreliable travel times. A major source of this unreliability is that when roads are congested, interactions between drivers can lead to capacity unexpectedly falling. For example, collisions can close lanes and aggressive lane changers can slow traffic. This paper analyzes how tolls should be set when accounting for such endogenous reliability. We find tolls should be higher and maximum flow lower than we might naïvely expect; and that such tolls make homogeneous drivers better off, even before the toll revenue is used. Simulations suggest the socially optimal maximum departure rate is 15% below that which maximizes expected throughput, and that tolling reduces private costs by almost 10%.