Estimating the statistics of events occurring around us is a crucial feature for survival. However, how the brain achieves this is poorly understood. We devised an odor-discrimination task in which mice could utilize preceding odor cues to estimate the probability of encountering a followed rewarded or non-rewarded odor. Recording the neural activity during task performance in the primary olfactory cortex, we found two non-overlapping neuronal subpopulations that encoded the odor-probability contingencies. One subset of neurons ramped its firing rate between the cue and the predicted odor onset proportionally to the estimated probability. The second subpopulation encoded the odor prediction error, with a more robust response to the rewarded odor when it was less probable and a weaker response when it was more probable. Reversing the probability contingencies remapped the activity in these neuronal subpopulations to reflect the newly learned probabilities. Bilaterally silencing the orbitofrontal cortex using DREADDs hampered mice ability to utilize the cue-probability contingencies and compromised neuronal probability coding in the olfactory cortex, while task learning remained intact. These results demonstrate that primary olfactory cortex neurons encode the probability of an odor event and its prediction error, and that these encodings require centrifugal inputs from the orbitofrontal cortex.