Top multimodal models have about 200-1000 tokens per image, so the math works out.