Theo says Codex hides the model from its own agent
A year of complaints later, the agent still only knows it is "based on GPT-6".
Theo says Codex still does not put the current model in its system prompt, which means the agent has no idea what it is beyond "based on GPT-6". He says he has complained about it for a year straight and that it has not been fixed for no good reason.
Your agent has no idea what it is beyond "based on GPT-6".
He tagged @thsottiaux asking whether the change could land in one of the 28 days of Codex improvements, and offered to give up a reset to get it done. That is a usage reset, the kind of credit that is normally worth hoarding, which is a reasonable proxy for how much the gap bothers him.
Why a missing string matters
For working engineers, this is a harness problem rather than a model problem. The system prompt is the text the tool injects before your conversation starts, and it is where an agent learns the things it cannot infer, including what it is running on. If the current model is absent, the agent is left guessing from whatever it absorbed in training, which is why it falls back to a generic "based on GPT-6" self description rather than naming the specific model serving the request.
That matters for anything where the right move depends on which model is answering. An agent that knows it is on a smaller, faster model might plan shorter steps or hand work off. An agent that thinks it is something else may make the opposite call. None of this shows up in model evaluations, because evaluations score the model, not the prompt the product wraps around it.
One caveat worth keeping. This is one engineer's complaint about one tool, not a measurement, and Theo does not report a specific task that failed because of the omission. What he does report is duration. A year of asking, same behavior, and an open offer to pay for the fix in usage.
Have the good opinion, not the loud one
One email a day. Under three minutes. What working engineers found out about coding with AI.