Robustness to image corruption is essential for the safe and reliable deployment of deep learning systems in safety-critical applications. While prior work has extensively investigated the corruption robustness of Convolutional Neural Networks (CNNs) and Vision Transformers (ViTs), the robustness of Multimodal Large Language Models (MLLMs) remains underexplored—particularly in direct comparison with vision-only architectures. To bridge this gap, we propose UniCorrupt-Bench, the first unified benchmark designed to systematically evaluate and compare the corruption robustness of CNNs, ViTs, and MLLMs under consistent settings. We conduct a comprehensive evaluation of nine representative models across a diverse range of common corruption types. Our analysis reveals that MLLMs generally exhibit superior robustness compared to traditional vision models, although they show slightly reduced resilience to noise and color distortions relative to ViTs. We observe that geometric transformations and color shifts have limited impact on performance, whereas compression artifacts and blurring present persistent challenges across all architectures. Notably, the robustness advantage of MLLMs appears to stem primarily from their capacity to extract higher-quality visual representations, rather than from the language components that distinguish them from vision-only models. These findings offer novel cross-architecture insights and emphasize the importance of robust feature extraction in multimodal systems. The full benchmark, appendix, and codes are publicly available at https://github.com/EdyQiu/UniCorrupt-Bench .