Recent advances in foundation models offer promising, scalable solutions for digital pathology, especially for classification tasks with limited data. However, most studies treat these models as static feature extractors, overlooking their full potential in real-world scenarios. Meanwhile, accurate blood cell morphology analysis is vital for diagnosing disorders like anemia and leukemia, particularly in resource-limited settings. In this study, we systematically evaluate the robustness and adaptability of four pathology foundation models specifically Gigapath, PLIP, DinoBloom, and PhikonV2, across five hematology datasets with diverse sizes, class distributions, and clinical contexts. We benchmark each foundation model under different training paradigms and compare them against standard deep learning architectures. Our results reveal that model rankings are highly dependent on the size of the training data and model configuration, with foundation models often underperforming or matching simpler models when not fine-tuned appropriately. These findings emphasize the need for robust, multi-regime benchmarking to fairly assess foundation models in pathology, especially in domain-specific, low-data applications like hematology.