Task-constrained deep learning models have shown strong performance for medical image segmentation. Recently, generalist segmentation foundation models have emerged, showing promising results across different scenarios. However, we lack large-scale studies comparing the performance of 2D image, video, and volume segmentation across multiple models and modalities. To unmask how foundational the models truly are, we comprehensively evaluate the segmentation performance of SAM2.1, SAM3, MedSAM2, SAM-Med2D, SAM-Med3D, nnInteractive and VISTA3D on more than 80 medical datasets. MedSAM2stands out as the most foundational of all models, while SAM-Med3D and VISTA3D excel in 3D CT segmentation scenarios, but require noticeably more computational power and memory. nnInteractive seems to be a promising model, featuring very fast inference time and rather high segmentation performance. Our code and evaluation results are openly available at https://github.com/DavidL-11/med-seg-fm.