“OpenAI has a strong bias towards using as few model architectures as possible to handle multiple modalities because maintaining and scaling separate infrastructure stacks is a significant, underestimated cost.”