[Windows][Codex Desktop] Possible silent model routing substitution: GPT-6 Astra returned GPT-5.6 Luna #47284

What version of the Codex App are you using (From “About Codex” dialog)?

Unknown (Codex desktop app; current build)

What subscription do you have?

ChatGPT account sign-in (plan not specified; not API key)

What platform is your computer?

Windows desktop

What issue are you seeing?

Summary

A local model-routing probe observed a possible silent substitution in the Codex desktop app: a request for gpt-6-astra returned gpt-5.6-luna three times. No usage-limit, downgrade, or Luna Reserve notice was visible.

This is a report of an observation, not a claim that the substitution is intentional.

Observed data

• Date: 2026-09-22 (Asia/Shanghai)
• Requested model: gpt-6-astra
• Returned model: gpt-5.6-luna
• Occurrences: 3, around 12:29 and 12:31
• Control observation: a separate request for gpt-5.6-terra returned gpt-5.6-terra
• Surface: Codex desktop app
• Authentication: ChatGPT account sign-in, not an API key
• No limit, downgrade, or Luna Reserve message was shown

The observation was recorded by the public codex-model-watch probe: https://github.com/ysh1112/codex-model-watch

What steps can reproduce the bug?

1. Open the Codex desktop app on Windows and sign in with a ChatGPT account.
2. Request or probe gpt-6-astra and record the model reported by the response or routing metadata.
3. Repeat the observation around the same time; the local probe recorded three gpt-6-astra requests returning gpt-5.6-luna.
4. Run the same check with gpt-5.6-terra as a control; it returned gpt-5.6-terra.
5. No usage-limit, downgrade, or Luna Reserve banner was present.

What is the expected behavior?

When gpt-6-astra is selected or requested, Codex should either return gpt-6-astra or clearly disclose the fallback model and the reason for it.

Additional information

Could the Codex team confirm whether this is expected routing, a capacity/safety fallback, a client display issue, or a bug? If fallback routing is expected, please document how users can detect it and whether an opt-out or explicit notification is available.

OpenAI Support has been contacted separately; no credentials, private logs, or local database have been attached to this issue. The exact desktop app build was not recorded in the local observation.