Prernack's probability hypothesis provides a simple empirical rule for predicting reinforcement effects, but has always been applied to response probabilities estimated by averaging over entire sessions. If the rule is robust, it should also predict momentary (e.g., withinsessions) changes in reinforcement from parallel momentary probability changes. It seems to do so. Six rats received noncontingent water (base), then leverpressed for water (contingency), each for 15 sessions. All sessions were divided into six subsessions. Average leverpressing for individual rats was a simple monotonic, usually linear, function of the probability of drinking-estimated from that subsession's counterpart during base. Similar results were obtained from a second study even though different instrumental and contingent events were used. With some generality, then, it is possible to apply the probability hypothesis to momentary reinforcement effects.A simple rule for predicting reinforcement has been proposed: if one response (A) is more probable than another response (B), A will reinforce B when made contingent upon B (premack, 1959, 1965). It has also been suggested that the magnitude of the reinforcement effect tends to be proportional to the magnitude of the probability difference between A and B (Premack, 1971 ;Terhune & Premack, 1974). Probability is a mathernatical idealization, but the rule or "probability hypothesis" was made operational by assuming that relative response durations would serve as reasonable estimates of response probabilities. Considerable evidence for the hypothesis has been obtained when response probabilities were estimated by the ratio of actual time spent in a response state to the time that could possibly be spent in the state (see Premack, 1971, for a review). Later, the hypothesis was extended to the punishment contingency (premack, 1971), and it was shown to hold in both the reinforcement and punishment situations when the probability estimate was the cumulative relative frequency of durations of the response (Terhune & Premack, 1970.In all previous investigations of the hypothesis, estimates of noncontingent response probabilities were obtained by averaging data over a block of base sessions during which the responses to be used in the contingency were not restricted and either one or both could