Repository navigation
Model selection reverts to gpt-6-astra in every new remote chat #47710
Description
What version of the Codex App are you using (From “About Codex” dialog)?
ChatGPT 26.917.62051
What subscription do you have?
ChatGPT Pro (20x)
What platform is your computer?
Apple M1 Pro, Darwin 27.0.0 arm64 arm
What issue are you seeing?
When I start a new chat in a remote project and select a model other than gpt-6-astra, the model picker briefly displays the selected model, then automatically reverts to gpt-6-astra.
This happens every time and prevents me from selecting another model for a new remote chat. The selection reverts before I send the first message.
Screen.Recording.2026-09-24.at.10.12.13.AM.mov
What steps can reproduce the bug?
- Connect to a remote environment using the desktop app's remote connection feature.
- Select a remote project.
- Start a new chat.
- Select a model other than
gpt-6-astra. - Wait briefly without sending a message.
- Observe that the model picker reverts to
gpt-6-astra.
Reproducibility: Every time.
What is the expected behavior?
The selected model should remain selected and be used when I send the first message. It should not automatically revert to gpt-6-astra.
Additional information
The remote environment is connected through the ChatGPT desktop app's remote connection feature.
I will attach a screen recording demonstrating the model selection reverting.
Activity
TheIgnacioLok commentedon Sep 24, 2026
Mismo problema en Windows
kostysh commentedon Oct 1, 2026
Also reproduced on Linux in a local project — model and reasoning both reset
I am seeing the same visible selection reset in Codex in the ChatGPT desktop app, in a local project, before sending the first message. Reasoning effort also resets. This adds a Linux/local-project reproduction to the original remote/macOS report; I am not claiming a shared root cause.
Environment: ChatGPT desktop 26.928.31416, Ubuntu 26.04.1 LTS, x86_64. Recorded on October 1, 2026, at approximately 17:49 and 17:54 Europe/Rome (UTC+2).
Steps and actual behavior
Expected: The explicit model and reasoning effort selected for the new chat should remain selected and apply to its first request. Defaults should supply the initial selection. If selection is restricted, the picker should explain that restriction instead of silently reverting an enabled control.
Evidence limits: These recordings demonstrate the visible selection reset. They do not establish which model or effort a submitted request would actually use. I have not verified a workaround or fix.
Related reports: #47859 (new-chat model and effort selection) and #44769 (effort resets before the first message).
Model selection reverting — 18.95 seconds
Screencast.From.2026-10-01.17-49-11.mp4
Reasoning effort reverting — 12.07 seconds
Screencast.From.2026-10-01.17-54-51.mp4
kostysh commentedon Oct 1, 2026
Update: verified workaround
I tested this on another system similar to the Linux setup in my earlier report. Commenting out both project-level model and reasoning defaults in the project's
.codex/config.toml, then restarting the ChatGPT desktop app, resolves the selection reset in that environment:After the restart, I can immediately select a different model and reasoning effort in a new project chat, before sending the first message. The selection no longer snaps back to GPT-6.1 Sol / Max.
Only these two defaults need to be commented out; the rest of the project configuration can remain in place.
This is a verified workaround, rather than an application fix. The tested combination was both defaults commented out + app restart; I have not tested the two settings individually or whether the restart is strictly necessary.
Follow-up to my earlier reproduction with screen recordings: #47710 (comment)