The authors introduce OCR-Robust, a benchmark designed to evaluate the robustness of vision-language models during OCR reasoning tasks under visual perturbations. The dataset comprises 812 samples divided into two subsets: OCR1.0, which covers documents and handwriting, and OCR2.0, focusing on charts and tables. A pilot study identified five representative perturbation types at three severity levels to ensure efficient evaluation. The study benchmarks 18 models, including proprietary systems and open-source VLMs, using metrics like Relative Corruption Retention and Worst-Case Retention. Results indicate that higher clean accuracy does not necessarily correlate with stronger robustness against visual degradation. Furthermore, the analysis reveals that charts and tables are substantially more fragile than document-like inputs when subjected to these perturbations.