Previous research has shown that individuals do not respond to changes in their bargaining position to the extent predicted by standard bargaining theories. Most of these results come from experiments with bargaining power allocated exogenously, so that individuals may perceive it as having been “unearned” and thus be reluctant to exploit it. Typically these experiments also allowed equal splits of the “cake” (the amount bargained over) as equilibrium outcomes, leading to a powerful tendency toward 50-50 splits. We conduct a bargaining experiment in which subjects earn their bargaining power through a real–effort task. Treatments are based on the Nash demand game (NDG) and an unstructured bargaining game (UBG). Subjects bargain over a fixed amount of money, with disagreement payments determined entirely by the number of units of the real–effort task successfully completed. Task parameters are set to allow disagreement payoffs above half the cake size, in which case 50–50 splits are not individually rational, and thus not consistent with equilibrium. We find that subjects are least responsive to changes in own and opponent disagreement payoffs in the NDG with both disagreement payments below half the cake size. Responsiveness is higher in the UBG, and in the NDG when one disagreement payment is more than half the cake size, but in both cases it is still less than predicted. It is only in the UBG when a disagreement payment is more than half the cake size that responsiveness to disagreement payoffs reaches the predicted level. Our results imply that even when real–life bargaining position is determined by past behaviour rather than luck, the extent to which actual bargaining corresponds to theoretical predictions will depend on (1) the institutions within which bargaining takes place, and (2) the distribution of bargaining power; in particular, whether the 50–50 norm is a viable outcome.