Many decisions benefit from the accumulation of evidence obtained sequentially over time. In such circumstances, the decision maker must balance speed against accuracy, and the nature of this tradeoff mediates competing desiderata and costs, especially those associated with the passage of time. A neural mechanism to achieve this balance is to accumulate evidence in suitable units and to terminate the deliberation when enough evidence has accrued. To accommodate time costs, it has been hypothesized that the criterion to terminate a decision may become lax as a function of time. Here we tested this hypothesis by manipulating the cost of time in a perceptual choice-reaction time task. Participants discriminated the direction of motion in a dynamic random-dot display, which varied in difficulty across trials. After each trial, they received feedback in the form of points based on whether they made a correct or erroneous choice. They were instructed to maximize their points per unit of time. Unbeknownst to the participants, halfway through the experiment, we increased the time pressure by canceling a small fraction of trials if they had not made a decision by a provisional deadline. Although the manipulation canceled less than 5% of trials, it induced the participants to make faster decisions while lowering their decision accuracy. The pattern of choices and reaction times were explained by bounded drift-diffusion. In all phases of the experiment, stopping bounds were found to decline as a function of time, consistent with the optimal solution, and this decline was exaggerated in response to the time-cost manipulation.