Function tools with reasoning_effort are not supported for gpt-5.4 in /v1/chat/completions

[SwAt1563]

pipecat version

0.0.105

Service Name

OpenAI

Service or model version

gpt-5.4

Issue Description

https://developers.openai.com/api/docs/guides/reasoning

Reproduction Steps

llm = OpenAILLMService(
        api_key=settings.llm.openai.OPENAI_API_KEY,
        settings=OpenAILLMSettings(
            model="gpt-5.4",
            extra={"reasoning_effort": "low"},
        ),
    )

Expected Behavior

Actual Behavior

Error Logs

2026-03-16 21:13:14.946 | ERROR    | pipecat.processors.frame_processor:push_error_frame:744 - OpenAILLMService#0 exception (/usr/local/lib/python3.13/site-packages/openai/_base_client.py:1669): Error during completion: Error code: 400 - {'error': {'message': 'Function tools with reasoning_effort are not supported for gpt-5.4 in /v1/chat/completions. Please use /v1/responses instead.', 'type': 'invalid_request_error', 'param': 'reasoning_effort', 'code': None}}

2026-03-16 21:13:14.946 | WARNING  | pipecat.pipeline.service_switcher:handle_error:196 - Service OpenAILLMService#0 reported an error: Error during completion: Error code: 400 - {'error': {'message': 'Function tools with reasoning_effort are not supported for gpt-5.4 in /v1/chat/completions. Please use /v1/responses instead.', 'type': 'invalid_request_error', 'param': 'reasoning_effort', 'code': None}}

[omChauhanDev]

Hey @SwAt1563, good catch. This is actually a new hard constraint from OpenAI, not a Pipecat bug.

For gpt-5.4, OpenAI now explicitly blocks combining reasoning_effort with function tools on the standard /v1/chat/completions endpoint. If you want to use that combo, they force a migration to their brand new /v1/responses API.

Since our OpenAILLMService is currently wired to the chat completions endpoint, passing both is throwing that 400 error.

Workaround for now: Just drop reasoning_effort from your extra settings if you need function tools to work with that model.

The fix: Shifting to /v1/responses is a heavy lift. It's a big architectural change bcz it completely alters the payload schema & how context/tool calls are structured. There's a community PR open exploring this (#3164), but the maintainer team will likely be tackling this migration internally very soon since it requires some deep, framework-level context aggregator reworks.

Appreciate you reporting this, we're on it!

[markbackman]

@omChauhanDev is right. We're working on adding the Response API in a new module. We'll have this soon enough!