Artificial intelligence (AI) is increasingly used in healthcare decision-making, where predictive models must accurately distinguish patient risk, provide reliable probabilities, and remain computationally feasible for routine clinical deployment. Despite the emergence of tabular foundation models, evaluation has focused predominantly on predictive accuracy, providing limited evidence of their practical advantages in healthcare. This study evaluates the tabular foundation model TabPFN and calibrated gradient-boosted decision tree (GBDT) ensembles across multiple publicly available clinical prediction datasets. The calibration comparison followed a pre-registration fixing the equivalence margin, an inclusion gate that removed models that had not learned, and an explicit falsifier. Energy was read from processor power counters, each figure tagged by hardware, configuration, and batch regime. TabPFN was better calibrated on all five datasets across four calibration metrics. A saturated dataset carries no discriminating weight, so the confirmatory conclusion rests on the other four. Rejecting the pre-registered equivalence hypothesis indicates a calibration difference favoring TabPFN, rather than equivalence, although the gap was small (mean adaptive expected calibration error = 0.011). TabPFN showed better discrimination on four of five datasets, with a mean AUROC difference of 0.013 and intervals that exclude zero. Decision-curve analysis showed a small positive net benefit at most thresholds. Its modest gains in calibration, discrimination, and net benefit came at substantially higher computational cost. Tree models were more energy-efficient beyond approximately six predictions; per-prediction cost jumped 121-fold with a single configuration flag, and a validated cost model showed inference growing steeply with context size, making CPU-only deployment impractical at clinical scale. Thus, TabPFN’s advantage is narrow and most justified when calibrated probabilities carry clinical weight, volume is low, and an accelerator is available, where cost is characterized by timing rather than measured power.