Agent is reading back some Retell state transition details randomly

We’ve been using an agent for close to a year live in prod with ChatGPT 4 and are trying out 5.2 yesterday. In the last 24 hours we found that we are getting a some random agent responses to the customer with behind the curtain state transition details. Not sure if this is related to 5.2 over 4 but have not seen this at all in the last 8 months and obviously we dont have anything in our prompts indicating it should expose any hidden workings.

Example from call log (call_2ae469709bfda0c932d62b03be7):

Agent 2:23

This location requires prepayment, so I’ll text you a secure payment link 
after we finish up.】【”】【assistant to=functions.transition_to_confirm_submit_order
 टिप्पणी  天天中彩票中 to=functions.transition_to_confirm_submit_order  尚度: 彩票招商ित्तി{}

Agent read all that to the customer.

Another call (call_bb88d44ef437ae5d454d4c68564)

Agent 0:08
{"tool_uses":[{"recipient_name":"functions.transition_to_delivery_setup","parameters":{}}]}
Sure we can do that but we'll text you a link. Would you like pickup or delivery?

Is this a bug or gpt 5.2 just over reaching?

Thank you for reaching out to Retell AI Support. We’ve received your ticket and our team will respond within 8 hours.

Hi @fred1

You can roll back to GPT 4 or 4.1 as they have more accuracy and give better answers, GPT-5.2 handles function-calling formatting differently than 4/4.1.

I still report your issue to the team with the call ids to look into

Best Regards.

It’s a frustrating suggestion. I found 5.2 handled state transitions better and overall the accuracy and answers were equivalent or better. And I am not talking about function calling formats. These are not tool response. These are your internal addition to control the state transitions that are being exposed in the output.

Hello @fred1,

The state transitions are not affected by the model you choose.
And it’s a known issue with GPT’s new models; GPT 4.1 is the most consistent right now.