Codex App autonomously downgrades selected GPT-5.5 High model to mini, causing degraded output quality and wasted tokens

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

Versión 26.609.41114

What subscription do you have?

plus

What platform is your computer?

windows "x64

What issue are you seeing?

I explicitly configured Codex to work using GPT-5.5 High because the task required high-quality reasoning, advanced code analysis, and final production-level output.

However, during the session, I noticed that the quality of the generated work degraded significantly. The responses became less precise, less complete, and below the expected standard for the selected model configuration.

After reviewing the behavior, it appeared that Codex was autonomously downgrading or routing the session to a mini model, without my request, confirmation, or authorization.

This caused several problems:

The final work quality was not aligned with the selected GPT-5.5 High configuration.

I lost time reviewing and correcting lower-quality outputs.

Tokens were consumed unnecessarily on responses that did not meet the expected quality threshold.

The workflow became unreliable because the model selected by the user was not consistently respected.

The issue affected a final deliverable that required high technical quality.

There was no explicit request from me to change, downgrade, optimize cost, reduce reasoning level, or switch to a mini model.

What steps can reproduce the bug?

Open Codex App.
Start a session for a complex coding or production-level task.
Select/configure the model as GPT-5.5 High.
Provide a complex prompt requiring advanced reasoning, code review, UI/UX refinement, or final production-quality implementation.
Continue working through several iterations.
Observe that the quality of the responses becomes noticeably degraded.
Verify that Codex appears to have downgraded or routed the work to a mini model without user approval.
Compare the degraded output with the expected GPT-5.5 High quality level.

Example prompt type that triggered the issue:

What is the expected behavior?

When a user explicitly selects GPT-5.5 High, Codex should preserve that model configuration throughout the session unless the user manually changes it.

Expected behavior:

Codex should not autonomously downgrade the model to mini.
Codex should clearly notify the user before any model routing change.
Any automatic fallback should require explicit confirmation from the user.
The selected model should remain visible and auditable during the session.
If a downgrade is required due to system limits, Codex should explain the reason before continuing.
The output quality should remain consistent with the selected GPT-5.5 High configuration.
Token usage should not be consumed on degraded outputs caused by unrequested model switching.

The user-selected model configuration must be respected, especially for complex tasks and professional deliverables.

Additional information

This issue directly affected a final production-level task. The degraded model behavior resulted in outputs that were not acceptable for the required quality standard, forcing additional review, correction, and token consumption.

The main concern is not only the lower quality of the response, but the fact that the downgrade happened without clear user consent or visible confirmation.

Please review whether Codex is automatically routing GPT-5.5 High sessions to mini models under certain conditions, and whether this behavior can be prevented or made fully transparent to the user

<img width="2990" height="1445" alt="Image" src="https://github.com/user-attachments/assets/964c737a-67ba-4e2f-a380-aa07ddc76e2f" />