Brain age prediction from T1-weighted MRI has emerged as a promising biomarker for detecting deviations from normative aging trajectories. A key challenge, however, is ensuring robustness across external datasets, a requirement for future clinical translation. We evaluated four pre-trained brain age prediction models (BrainAgeNeXt, Pyment, DeepBrainNet and ENIGMA) across four external datasets: ADNI, a local cohort including long-COVID patients and controls, and two healthy cohorts (JUK, RRIB) from the OpenNeuro platform, totaling 1,634 subjects. Models based on 3D convolutional neural networks showed lower bias (MAE = 3.7 and 4.6 years vs. 6.2 and 12.3 years) and variance (ASTD = 2.9 and 3.7 years vs. 4.7 and 8.9 years), with performance improving with larger and more diverse training sets. A regression-to-the-mean effect in brain age prediction was observed across all models, with strong age underestimation in ADNI controls. Within datasets, individuals with neurodegenerative conditions consistently exhibited older brain age compared to controls. Finally, analysis on explainability maps highlighted voxels from the lateral ventricles as the most influential in model predictions, consistent with previous reports.