Brain tumors are life-threatening conditions characterized by the uncontrolled proliferation of abnormal cells. Accurate and timely detection is essential, as delayed diagnosis may result in severe outcomes. Magnetic Resonance Imaging (MRI) remains the most widely used modality for brain examination due to its superior soft tissue contrast and resolution. Over the last decade, research in brain MRI analysis has evolved from conventional machine learning approaches to advanced deep learning frameworks. In this work, we examine the potential of Vision Transformer (ViT) and Swin Transformer architectures for brain tumor classification. The ViT model applies transformer mechanisms to image patches, enabling global contextual learning without convolutional operations. In contrast, the Swin Transformer introduces a hierarchical representation with shiftedwindow self-attention, allowing scalable and locality-aware feature extraction. We propose a hybrid strategy that integrates features from both pre-trained ViT and Swin models and evaluates them with eight classical machine learning classifiers. Experiments were conducted on the publicly available Br35H: Brain Tumor Detection 2020 dataset to assess classification accuracy and computational performance under limited data conditions. The results are validated using multiple evaluation metrics to provide a comprehensive performance analysis.