Twelve adult human subjects were exposed to a sunk‐cost procedure with two options: a mixed‐ratio schedule of points later exchangeable for money, and an escape schedule that cancelled the current trial and initiated a new one. The mixed ratio included four values, arranged probabilistically in such a way that the expected ratios favored either persistence or escape. These probabilities were varied systematically on a within‐subject basis across conditions. Absolute ratio size was thus varied across four groups of three subjects each, yielding unique combinations of expected ratios from escaping and persisting. When the differences between escaping and persisting differed the least, subjects tended to persist, committing the sunk‐cost error. When the differences between persisting and escaping differed by a larger margin, choice patterns tended toward optimal—escaping or persisting as a function of the contingencies. These findings demonstrate that sunk‐cost decision‐making errors in humans are sensitive to their relative costs and benefits, and illustrate a promising set of methods for bringing such behavior under experimental control in the laboratory.