As token costs for LLMs decline toward zero, fundamental models increasingly look like commoditized technology. How can OpenAI and its competitors avoid this trap?