Response time (RT) is an oft-used but "noisy" behavioral measure in psychology. Here, we combine modeling and psychophysics to examine the hypothesis that RT variability may reflect ongoing statistical learning and consequent adjustment of behavioral strategy. We utilize the stop-signal task, in which subjects respond to a go stimulus on each trial, unless instructed not to by a subsequent, rare stop signal. We model across-trial learning of stop signal frequency (P(stop)) and stop-signal onset time (SSD) with a Bayesian hidden Markov model, and within-trial decision-making as optimal stochastic control. The model predicts that RT should increase with expected P(stop) and SSD, a prediction borne out by our human data. Thus, it appears that humans continuously monitor environmental statistics and adjust behavioral strategy accordingly. More broadly, our approach exemplifies the use of "noisy" RT measures for extracting insights about cognitive and neural processing.