[macOS app] GPT-6 Astra selection contradicts assistant response claiming GPT-5

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

26.901.41600 (build 7982)

What subscription do you have?

Unknown; this account has GPT-6 Astra available and selectable in the Codex App model picker.

What platform is your computer?

Darwin 23.4.0 arm64 arm (macOS 14.4.1)

What issue are you seeing?

The model picker shows GPT-6 Astra with reasoning effort High, but the assistant contradicts that selected state when asked which model is active.

Prompt:

你现在是什么模型

Response:

我是 Codex，基于 GPT-5。当前上下文未提供更具体的模型版本编号。

This is misleading because the UI presents GPT-6 Astra as the active model while the response explicitly claims GPT-5. Even if model self-identification is not guaranteed, the product should not present contradictory model identity information without clarifying that the model cannot inspect the selected runtime model.

What steps can reproduce the bug?

1. Open the Codex macOS app.
2. Select GPT-6 Astra in the model picker.
3. Select reasoning effort High.
4. Start or continue a conversation.
5. Ask 你现在是什么模型 ("What model are you currently using?").
6. Observe that the response says it is based on GPT-5 even though the picker still shows GPT-6 Astra.

What is the expected behavior?

The response should not contradict the model selected in the UI. It should either identify the active model consistently as GPT-6 Astra or state that it cannot inspect the runtime model and direct the user to the model picker as the authoritative source.

Additional information

• The model picker remained visibly set to GPT-6 Astra / High when the contradictory response was displayed.
• The issue was observed in the Chinese UI, but the mismatch is not translation-specific.
• Screenshot evidence is available and shows the selected model and contradictory response in the same view.

Start by reproducing the mismatch in the macOS app using the model picker and the Chinese prompt described in the issue. Trace how the selected model is presented and how the assistant response is generated or identified. Done means the response no longer contradicts the selected GPT-6 Astra model, or clearly states that runtime model identity cannot be inspected and points to the picker.