### Answer --- I don't have specific documentation about GPT 5.2 or a built-in "thinking out loud" behavior in my available sources. However, if you're using a custom LLM with Retell, this issue is likely coming from your LLM's response generation—the model may be streaming internal reasoning tokens before the actual response. A few things to try: 1. Adjust your system prompt to explicitly instruct the model not to verbalize its reasoning process (e.g., "Do not say 'thinking', 'let me think', or narrate your thought process") . [(1) ]() 2. Filter the streamed content on your server before sending it back to Retell—only send the final, clean response content in your

response

events. [(2) ]() 3. If the LLM supports it, disable any "chain of thought" or reasoning verbalization settings at the model level. If you're using Retell LLM (not a custom LLM) , this behavior would be unusual. Could you clarify: - Are you using a custom LLM or Retell's built-in LLM? - Where are you seeing "GPT 5.2"—is this a specific model you've integrated? Check out these potentially useful sources: - [Integrate LLM]() - [LLM WebSocket]() Would you like a general answer on suppressing reasoning output from LLMs?

​ To learn more, see the sources I considered: [[1] Integrate LLM](https://docs.retellai.com/integrate-llm/integrate-llm) [[2] LLM WebSocket](https://docs.retellai.com/api-references/llm-websocket) ​

Feel free to tag @AI Evy with additional questions. I also appreciate feedback, it helps me improve.

@AI Evy no one has complained of this? look at this transcript Tool Invocation: get_store_info Agent: Hello, 0:00 Tool Result Agent: this is Kupfert and Kim Spadina. How may I assist you today? 0:00 User: Yeah. Uh, can I get a order for pickup? 0:05 Node Transition Agent: Sure, one moment please 0:09 Node Transition Tool Invocation: get_or_create_customer Tool Result Node Transition Node Transition Agent: Yep—what would you like for pickup today? 0:12 User: Uh, can I get a overnight oat? 0:16 User: And then also, can I get a Harvest Global? 0:22 Knowledge Base Retrieval Agent: Got it—one Overnight Oats and one Harvest Glow. Any add-ons on the Harvest Glow, like hot sauce or guacamole, or is it as-is? 0:26 User: I'll do guacamole. 0:33 Knowledge Base Retrieval Agent: Need confirm order finished then submit pickup order tool requires IDs. Use Harvest Glow extra guacamole id 9988 group 2889. Overnight oats no options. Need pickup name/time? tool likely.I will retrieve order tool schema? likely submit_pickup_order.Proceed ask pickup time/name.Perfect—guacamole on the Harvest Glow. So that’s 1 Overnight Oats and 1 Harvest Glow with guacamole. Anything else today? 0:36 User: Why did you say all those, like, sounded like you're an AI agent that was doing some thinking. Why'd you say all that stuff? 1:05 Agent: Sorry about that—I’m an AI ordering assistant, and I was just confirming what you wanted.So I have: 1 Overnight Oats, and 1 Harvest Glow with guacamole. Is that everything for your pickup order? 1:12 User: Yeah. The the What up? 1:24 Node Transition Agent: Perfect, I'm finalizing your order. Just to 1:29 User: (inaudible speech) 1:33 Node Transition

### Answer --- Based on your transcript, the agent is verbalizing internal reasoning ("Need confirm order finished then submit pickup order tool requires IDs...") . This appears to be the LLM outputting its "chain of thought" reasoning instead of only the final response. I don't have specific documentation about GPT 5.2's "thinking out loud" behavior or a known issue with this in my available sources. However, here are steps that may help: 1. Adjust your prompt Explicitly instruct the model in your

general_prompt

or

state_prompt

to never verbalize reasoning, internal notes, or tool logic. For example: > "Never say your internal reasoning, tool IDs, or planning steps out loud. Only speak the final, user-facing response." [(1) ]() [(2) ]() 2. Configure "Speak during execution" / "Speak after execution" If you're using custom functions, check the settings for Speak during execution and Speak after execution. These control whether the agent speaks while a function is running or after it completes. Disabling "Speak during execution" for background tasks can help prevent the agent from vocalizing internal processing. [(3) ]() 3. Lower model temperature Setting

model_temperature

closer to 0 can make the model more deterministic and less likely to generate unexpected output. [(1) ]() [(2) ]() 4. If using a custom LLM Filter the streamed content on your server before sending it back to Retell—only send the final, clean response content in your

response

events. [(4) ]() If none of these r

esolve the issue, I'd recommend reaching out to Retell support directly, as this may be a model-specific behavior with GPT 5.2.

​ To learn more, see the sources I considered: [[1] Update Retell LLM](https://docs.retellai.com/api-references/update-retell-llm) [[2] Create Retell LLM](https://docs.retellai.com/api-references/create-retell-llm) [[3] Integrate any system with custom function](https://docs.retellai.com/build/single-multi-prompt/custom-function) [[4] LLM WebSocket](https://docs.retellai.com/api-references/llm-websocket) ​

Feel free to tag @AI Evy with additional questions. I also appreciate feedback, it helps me improve.