The brains’ ability to predict sensory input and infer its sources guides a wide range of cognitive processes. While predictive processing clearly optimizes informational aspects of perception and decision making, a biologically realistic implementation must take the energetic limits of brain metabolism into account. Interestingly, theoretical and empirical studies suggest that minimizing the mismatch between internally generated predictions and bottom-up input also minimizes energetic cost. However, the relationship between conventional imaging techniques like BOLD and EEG and metabolic processes is complex, indirect, and difficult to obtain on a systems level. To address this methodological shortcoming, we here apply a novel multiparametric MR method that quantifies the cerebral metabolic rate of oxygen on a voxel level during visual stimulation with varying levels of predictability. Using a statistical learning paradigm, we show that oxygen use is affected throughout large networks, extending beyond local peaks of activity in functionally specialized regions. Furthermore, interindividual variability in prediction error cost scales with subjective confidence regarding learning performance. The resulting metabolic differences exceed five percent on a network and whole-brain level - an effect magnitude usually observed only locally or between different tasks. In summary, our quantitative MR method shows that valid predictions are central to the brain wide energy balance as large amounts of resources are allocated to process deviations. Finally, the explanatory power of confidence ratings shows that cognitive energy expenditure can be derived from metacognitive judgements that fit within a Bayesian framework.