Is there a way to use the GPT-5.2 Pro model directly in Codex CLI?
Open 💬 7 comments Opened Jan 28, 2026 by Shuvomoy
💡 Likely answer: A maintainer (github-actions[bot], contributor)
responded on this thread — see the highlighted reply below.
What feature would you like to see?
I recently subscribed to ChatGPT Pro plan, and for my specific coding tasks, the GPT-5.2 Pro model on the web interface performs very well. I have also tried the Codex CLI for the same problems, but the quality of its output is not as good as that of GPT-5.2 Pro.
I am unsure if there is a way to access the specific GPT-5.2 Pro model within the Codex CLI. The CLI offers gpt-5.2 and gpt-5.2-codex with reasoning levels ranging from Low to Extra High; however, I have found that even the 'Extra High' setting on these models is inferior to the web-based GPT-5.2 Pro for my use case.
Any tips/suggestions will be much appreciated.
Additional information
OS: Ubuntu
ChatGPT Plan: Pro
7 Comments
Potential duplicates detected. Please review them and close your issue if it is a duplicate.
Powered by Codex Action
@Shuvomoy there isn't - but you could create a custom _skill_ that creates a
context.zipand aprompt.mdfor you to share with GPT 5.2 Pro via the ChatGPT app / interface.The GPT 5.2 pro model isn't recommended for coding use cases. It uses a "best of n" mechanism to produce results.
If you want high-reasoning code generation, we recommend using gpt-5.2 or gpt-5.2-codex with "xhigh" reasoning effort.
Thanks so much for the tips @guidedways this is very useful. I will try that!
Okay, I will try that, thanks @etraut-openai !
I think we’re underestimating how powerful Codex would become if it could call GPT-Pro as a sub-agent.
That unlocks real Build Review loops, where Codex orchestrates and Pro plays the heavyweight designer or adversarial reviewer.
Right now I manually simulate this:
The bottleneck is me. I’m mostly just shuttling outputs between sessions.
If Pro were callable as a skill inside a sub-agent, Codex could run this loop autonomously with explicit convergence or starvation criteria. You’d kick off a research loop, walk away, and come back to a refined, adversarially stress-tested design.
Think of this not as a coding mode, but more like a supercharged Plan mode, where codex will be able to design an entire new product from scratch.
Normally it's true, but pro can be useful in some case like program relies on mathematical strategies or requires very complex algorithm.
Like @vitarb said, calling pro model with collab(subagent) may be very useful. We don't need gpt-5.3-pro-codex. We just want to use existing pro model as subagent for planner, designer, etc.