Because GPT-4 is multimodal, which puts it into the entirely different class. GPT-3.5 was trained on text, same as this model.