Why OpenAI Models Struggle with PDF Extraction(and Why Gemini Fairs Much Better)

Does this only apply to PDFs with many images? If it is mostly text and table, could it just extract plain text?