[Linux] Regular Chat reasoning resets/sticks to Instant while Work and Web reasoning continue to work

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

26.901.31953 (chatgpt-26.901.31953-1.x86_64)

Bundled Codex binary is shipped inside the desktop package.

What subscription do you have?

ChatGPT Pro 5x (~€114/month).

What platform is your computer?

Fedora Linux 44, x86_64, KDE Plasma, Wayland.

What issue are you seeing?

In regular Chat in the unified ChatGPT/Codex desktop app, changing the reasoning level does not persist correctly. The UI falls back to / remains on Instant.

The issue is isolated to the desktop Chat surface:

• Desktop Chat: reasoning falls back to / displays Instant.
• Desktop Work: higher reasoning levels continue to work normally and have worked throughout the incident.
• chatgpt.com Web: higher reasoning levels continue to work normally with the exact same account.

This is not an intermittent issue. Work reasoning did not “start working again” after troubleshooting; it never stopped working. Web reasoning also remained functional throughout.

There is also a contradictory UI state in Desktop Chat. For example, after selecting Pro / GPT-5.5, the UI can remain visually in the Pro state while simultaneously showing 5.5 Instant.

I am not claiming that the UI alone proves which backend model executed a request; the user-visible state itself is inconsistent and differs from Work/Web on the same account.

Usage-limit control

The relevant visible allowance had already fully reset and showed 100% available before reproducing the issue again.

After that reset:

• Web reasoning still worked.
• Desktop Work reasoning still worked.
• Regular Desktop Chat still fell back to Instant.

So a simple account-wide reasoning/model limit does not explain the observed surface-specific behavior.

What steps can reproduce the bug?

1. Sign in to ChatGPT/Codex Desktop on Fedora Linux.
2. Open a brand-new regular Chat conversation (not Work).
3. Open the model/reasoning picker.
4. Select a non-Instant reasoning level (for example High / Very High, depending on model availability).
5. Close/reopen the picker or observe the composer state.
6. The regular Chat surface falls back to / displays Instant.
7. Switch to Work in the same desktop application and select a higher reasoning level.
8. Observe that Work retains the higher reasoning setting normally.
9. Open chatgpt.com with the same account and observe that higher reasoning levels also work normally there.

Reproduction rate: consistent.

Troubleshooting already completed

The issue persisted after all of the following:

• complete removal and fresh reinstall of ChatGPT Desktop;
• fresh zero-state profile;
• no reuse of previous ~/.codex state;
• no reuse of previous ~/.config/Codex state;
• no reuse of previous cache/auth/session state;
• reproduction using a fresh stock profile;
• reproduction outside my custom bubblewrap/systemd isolation;
• sign-out / sign-in;
• reproduction again in a brand-new regular Desktop Chat;
• visible allowance fully reset to 100%;
• package integrity verification (rpm -V chatgpt exit code 0).

The custom isolation environment is therefore not a plausible cause of the original reasoning bug.

Diagnostic evidence

A timestamp-aligned reproduction was recorded at approximately 2026-09-04 18:52:24 CEST.

Relevant desktop log event during the same capture:

2026-09-04T18:52:45.176229+02:00
Setting default model and reasoning effort newEffort=high newModel=gpt-5.6-sol

This corresponds to the correctly functioning Work behavior shown in the recording.

The same reproduction also contains calls such as:

configRequirements/read
modelProvider/capabilities/read
model/list

The regular Desktop Chat failure is primarily visible in the UI; I did not find an equally explicit log line saying that the UI had fallen back to Instant.

Feedback ID

Feedback was uploaded immediately after reproducing the issue in a brand-new regular Desktop Chat, as requested by Support.

Feedback ID:

no-active-thread-01a06d71-ec8c-70b2-910e-160973ebcc31

The app itself displayed “Feedback uploaded” and instructed me to mention this ID in an existing/public issue.

Expected behavior

Regular Desktop Chat should retain and apply the selected reasoning level, matching Desktop Work and chatgpt.com for the same account.

A reset of one model-specific quota should not force unrelated available reasoning levels to Instant if those same levels remain available on other surfaces.

Actual behavior

Regular Desktop Chat remains stuck on / falls back to Instant, while Desktop Work and Web continue to allow higher reasoning normally.

Related issues

This looks closely related to:

• #42435 — Windows app reasoning effort resets from Extra High to Instant. That report also shows Desktop resetting to Instant while chatgpt.com works correctly with the same account.
• #42531 — macOS Chat mode stuck on Instant while Work and other clients retain reasoning choices.

Because this report reproduces the same class of failure on Linux/Fedora, it may share the same desktop Chat model/reasoning-state handling code, although I am not claiming the root cause is necessarily identical.

Additional information

I can provide a short screen recording, sanitized desktop logs, and environment details. The full unfiltered journal contains more diagnostic data than necessary and is being kept private unless maintainers request it.

Public-report privacy: account identifiers, email addresses, conversation IDs, request IDs, tokens, cookies, and unrelated local paths have been omitted/redacted.

Start by reproducing the issue on Fedora Linux and comparing reasoning-state behavior between regular Desktop Chat, Work, and chatgpt.com. Review related issues #42435 and #42531 for shared desktop Chat handling clues. Done means regular Chat retains and applies a selected non-Instant reasoning level consistently with Work and Web.