Possible routing bug in GPT-5.2 Pro?

When I made a second turn with the model within the same conversation in the ChatGPT interface, asking it to edit the code it had generated, it replied:

“To apply these adjustments to the response/code you requested, you need to click the Update button and include exactly these changes in the task prompt so that the other model (which is running this task in ‘pro mode’) can regenerate the output with the modifications.

Once you click Update and resubmit with those three points, the full code will come back updated as you described.”

This response makes no sense in relation to what I asked. I simply asked the model to tweak some details in a Python code snippet, and it replied with weird things like “you need to click the Update button,” which has nothing to do with my code. It sounds like internal chatter about tool usage behind the scenes. Note that this weird response sounds like one model talking to another model. I was using GPT-5.2 Pro. Apparently, this reveals that the ChatGPT interface performs some kind of prior model selection/routing even when GPT-5.2 Pro is already selected. This seems like a bug.

can you take a screenshot?

Here it is, but it’s in Portuguese. Note that it thought for about 6 minutes and, instead of providing the answer, it returned this strange message. When I ran it again on another attempt, it worked without issues.

I also noticed that even when selecting the thinking model, sometimes it responds without thinking, apparently using the instant model. In other words, I strongly suspect that there is some routing going on behind every interaction.

yeah
I think so… Same situation in Codex CLI