Suspected hidden model substitution: selected 5.5 behaves like 5.6 and runs tasks endlessly without results

What issue are you seeing?

I want to report a serious regression in Codex/ChatGPT behavior.

The UI shows that I have selected 5.5. I choose this model intentionally because it previously worked better for my real Codex tasks: it understood the task faster, followed instructions more reliably, and completed work instead of circling around it.

Now the behavior of the selected 5.5 has changed sharply and feels like 5.6: the model works for a very long time, goes in circles, performs many actions, but does not bring the task to a finished result.

I cannot technically prove that the backend actually substituted the model. However, from the user side, this looks like a hidden effective-model change: the interface still shows 5.5, but the actual behavior feels like 5.6.

Observed behavior

• Tasks take much longer than before.
• The model can work for hours without producing a useful final result.
• It often goes in circles.
• It performs many steps but does not actually finish the task.
• It follows the original instructions worse than before.
• It has more trouble distinguishing important work from secondary work.
• The final working result is often missing.
• The UI still shows that 5.5 is selected.

Why this is critical

Model selection needs to be honest and predictable. If a user selects 5.5 because it works better for their tasks, the system should not silently move them to behavior that feels like a different model.

Even if this is not an intentional substitution and is instead caused by fallback, routing, capacity handling, or an internal model version change, the user should be able to see what effective model or mode is actually being used.

Otherwise, it becomes impossible to trust the model selector: the interface shows one thing, but the model behavior feels like something else.

Expected behavior

If 5.5 is selected, it should behave like 5.5: reliably execute Codex tasks, follow instructions, and finish the work.

If the system uses 5.6 or another effective model instead of the selected 5.5 for any reason, that should be clearly shown to the user.

Actual behavior

The selected 5.5 behaves like 5.6: it works for a long time, goes in circles, can spend hours on a task, and still may not reach a useful result.

Request

Please check whether:

• routing changed for users who selected 5.5;
• fallback to 5.6 is being used;
• the internal model version changed under the same visible model name;
• there is a capacity mode where the selected model is effectively replaced by another one;
• the actual/effective model can be shown to the user for each Codex thread.

The main issue is trust: the user needs to understand which model is actually executing the task. Right now that trust feels broken.