A response reproduction paradigm for investigating numerical ability in animals is presented. In this experiment subjects were presented a number of light flashes (2, 4, or 6) in a sample phase, and were then required to reproduce that number in a production phase by pecking one key either 2, 4 or 6 times and then pecking another key once. Flashes in the sample phase were presented in one of two manners. Either one flash occurred every 2.5-seconds (termed the rate-controlled procedure), or all flashes occurred within a 10-second interval (termed the time-controlled procedure). Two subjects were trained in each procedure. Following training, testing with novel probe trials (1, 3, 5 and 7-flash sequences) was conducted. Novel probe trials, interspersed with baseline trials, were either presented in the same manner as the baseline trials, consistent transfer, or in the opposite procedure to baseline trials, inconsistent transfer. Reinforcement was delivered in a random manner on probe trials, and only for correct responses on baseline trials. In a second condition training and testing conditions were reversed, so that subjects that had received time-controlled training now received rate-controlled training and vice versa. All subjects learned to respond in the task with reasonable accuracy. Response distributions for the three baseline trials were distinct from one another and peaked at, or near, the reinforced number of responses. The introduction of novel probe trials disrupted baseline performance to varying degrees. In consistent transfer testing number of responses increased as flash number increased in an orderly manner for all subjects. During inconsistent transfer testing differences between rate and time-controlled performance emerged. All subjects maintained to some degree their baseline trial performance, however, time controlled subjects showed little discrimination between probe trials, while rate-controlled subjects made fewer responses to trials consisting of large numbers off lashes (5 and 7) than to trials consisting of small numbers of flashes (1 and 3). Examination of standard deviations of responses and coefficients of variation suggest that subjects in the rate-controlled procedure were relying upon temporal cues. Subjects Q5 and Q8 in the time-controlled procedure also appear to have used temporal cues. A confusion-diffusion model describing subjects' performance in this experiment is also presented. The usefulness of the response reproduction paradigm for investigating animal numerical ability is discussed and modifications to the present procedure are suggested.