The ability to detect, represent and exploit statistical regularities in the environment is a key feature of animal cognition. The brain holds internal models of the world to optimally interpret incoming sensory inputs and guide the appropriate responses. Previously, we have shown that humans and rodents are sensitive to the underlying sensory statistics in a 2-alternative-forced-choice auditory categorisation task. In this work, we investigate the neural signatures of behavioural adaptation to changes in sensory statistics, where the sampling distribution of sounds within each category is modulated to create distinct “statistical contexts”. Switching from one context to another happens without cues, within and across sessions. Depending on the underlying statistics, subjects systematically change their psychometric performance. Using a normative model we show that such context-dependent adaptation is optimal for reward maximisation, given perceptual noise constraints. As animals flexibly adapt their choice strategy to different sensory statistics, we expect the neural representation of stimulus-to-action mapping to vary accordingly. We perform neural recordings with neuropixels in rats in medial prefrontal cortex (mPFC) and identify features of statistical context-dependent encoding of task variables at single cell and population levels. We show that single neurons are tuned to sensory stimuli in a statistical context-dependent manner. The population activity aligned to the choice axis is reorganised, contingent on the animal’s behavioural adaptation to the underlying context. In summary, our study shows that population activity in mPFC contains signatures of knowledge about the underlying sensory statistics, in service of the optimal adaptation of animals’ choice biases.