Voice mode, model-setting persistence, and inconsistent model selectors in the macOS app

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

26.721.41059

What subscription do you have?

ChatGPT Plus

What platform is your computer?

Darwin 25.5.0 arm64 arm

What issue are you seeing?

Title: Voice mode, model-setting persistence, and inconsistent model selectors in the macOS app

Application: ChatGPT — Powered by Codex & OWL
Version: 26.721.41059
Release date: July 24, 2026
Platform: macOS
Account: ChatGPT Plus

I am encountering three related issues involving Projects, voice mode, and model selection.

1. New voice mode is unavailable inside Projects

In the macOS application, I can use the new ChatGPT voice mode in a new chat created outside a Project. However, voice mode is not available in chats inside a Project.

The same account can use the new voice mode inside the same Project through the ChatGPT web application. This suggests that voice within Projects is intended to be supported but is not functioning correctly in the macOS app.

Expected behavior: Voice mode should be available inside Projects in the macOS app, as it is in the web app.

Actual behavior: Voice mode is available only in chats outside Projects.

2. Switching from voice back to text changes the selected reasoning level to Instant

For a complex, ongoing Project, I normally select ChatGPT 5.6 Sol at High or Extra High reasoning. I expect that selection to remain in effect until I explicitly change it.

Because voice mode was unavailable inside the Project in the macOS app, I opened the Project in the web app and used voice mode there. After providing extensive voice feedback about work in progress, I switched back to text chat. The model selector had changed to Instant without my noticing.

The resulting response was long but contained several errors and questionable judgments. After noticing the change, I manually selected High again and asked ChatGPT to review the Instant response. It identified multiple problems and substantially rewrote the answer.

This creates a serious reliability risk in complex work: changing the interaction modality can silently reduce the reasoning level immediately before the model processes important instructions.

Expected behavior: A user-selected reasoning level such as High should remain selected when switching between voice and text, unless the user explicitly changes it. If voice mode requires a different internal model configuration, returning to text should restore the user’s prior selection.

Actual behavior: Switching from voice mode back to text silently changed the selector to Instant.

3. Chat and Work use inconsistent model-selection systems

The model-selection options in Chat—such as Instant, Medium, and High—are different from the options presented in Work mode. This is confusing when Chat and Work are being used together on the same complex Project.

It is not clear whether similarly named or differently named settings provide comparable reasoning capability, whether a preference carries between modes, or how to ensure that both Chat and Work consistently use High or Extra High reasoning.

Suggested improvement: Use consistent reasoning-level terminology across Chat and Work, or clearly explain how the options correspond. Ideally, users should be able to set a persistent Project-level reasoning preference that applies across text, voice, Chat, and Work.

These issues are particularly consequential for users who rely on Projects for complex, iterative work involving multiple chats and Codex. Model choice should be predictable, persistent, and visible, and changing interfaces or input modes should not silently lower the selected reasoning level.

What steps can reproduce the bug?

Feedback ID: no-active-thread-019fa128-e314-7073-a314-7ad2ab4511f6

What is the expected behavior?

see report

Additional information

_No response_