More importantly to me, I think generating synthetic data is OpenAI's secret sauce (no evidence I am aware of), and they need access to GPT-4 weights to train GPT-5.