The "GPT in a wrapper" framing is exactly right.. If any sufficiently capable model can be the engine, the value isn't the model, it's the orchestration and context management around it.

Codex is a product bet that OpenAI can bundle faster than others can plug in competing models.