`OpenAIChatModel` ignores settings from `OpenAIResponsesModelSettings`

[duarteocarmo]

Initial Checks

• I confirm that I'm using the latest version of Pydantic AI
• I confirm that I searched for my issue in https://github.com/pydantic/pydantic-ai/issues before opening this issue

Description

When using OpenAIResponsesModelSettings with OpenAIChatModel they get completely ignored.

Just run examples with uv run

For example, when I run it, I get this:

Error in test: status_code: 400, model_name: gpt-5-nano, body: {'message': 'Your organization must be verified to generate reasoning summaries. Please go to: https://platform.openai.com/settings/organization/general and click on Verify Organization. If you just verified, it can take up to 15 minutes for access to propagate.', 'type': 'invalid_request_error', 'param': 'reasoning.summary', 'code': 'unsupported_value'}
ModelResponse(parts=[TextPart(content='Based on the latest official data, the city of Ancona (municipality) has a population a bit over 100,000, with the trend being fairly stable to very slowly increasing in recent years. A reasonable 2025 estimate would be about 102,000 people, give or take a few thousand.\n\nNotes:\n- This is an approximation; official numbers come from Istat (Italian National Institute of Statistics) and local records.\n- If you need a precise figure for a specific date in 2025, check Istat’s annual estimates or the Comune di Ancona’s population reports.')], usage=RequestUsage(input_tokens=19, output_tokens=321, details={'accepted_prediction_tokens': 0, 'audio_tokens': 0, 'reasoning_tokens': 192, 'rejected_prediction_tokens': 0}), model_name='gpt-5-nano-2025-08-07', timestamp=datetime.datetime(2025, 11, 20, 16, 44, 20, tzinfo=TzInfo(0)), provider_name='openai', provider_details={'finish_reason': 'stop'}, provider_response_id='chatcmpl-ceaceacac', finish_reason='stop')

Example Code

# /// script
# dependencies = [
#   "pydantic-ai",
# ]
# ///

from pydantic_ai import ModelRequest
from pydantic_ai.direct import model_request_sync
from pydantic_ai.models.openai import (
    OpenAIChatModel,
    OpenAIResponsesModel,
    OpenAIResponsesModelSettings,
)


def test():
    settings = OpenAIResponsesModelSettings(
        openai_reasoning_effort="low",
        openai_reasoning_summary="concise",
    )

    model = OpenAIResponsesModel("gpt-5-nano", settings=settings)

    model_response = model_request_sync(
        model=model,
        messages=[
            ModelRequest.user_text_prompt(
                "Population of city of Ancona in 2025? Estimate."
            )
        ],
    )
    print(model_response)


def very_weird_case():
    settings = OpenAIResponsesModelSettings(
        # Both of these get completely ignored?
        openai_reasoning_effort="low",
        openai_reasoning_summary="concise",
    )

    model = OpenAIChatModel("gpt-5-nano", settings=settings)

    model_response = model_request_sync(
        model,
        messages=[
            ModelRequest.user_text_prompt(
                "Population of city of Ancona in 2025? Estimate."
            )
        ],
    )
    print(model_response)


if __name__ == "__main__":
    for function in [test, very_weird_case]:
        try:
            function()
        except Exception as e:
            print(f"Error in {function.__name__}: {e}")

Python, Pydantic AI & LLM client version

Latest.

[DouweM]

@duarteocarmo That's expected/intentional: OpenAIChatModelSettings maps to OpenAIChatModel, OpenAIResponsesModelSettings maps to OpenAIResponsesModel, so if you want to configure the chat API, you can't expect it to accept settings made for the responses API

[duarteocarmo]

@DouweM - great. Thought would be better to error out instead of silently accepting, since I have just seen this in prod. Happy to close though.

[DouweM]

@duarteocarmo Model classes intentionally ignore unknown keys in the settings they're provided so that you can define settings for multiple models on the agent and then use it with different models passed in at runtime without hitting errors or having to do your own filtering. But yeah it has the downside of making it a bit harder to detect why a certain setting isn't "sticking"...

Let me know if you have an idea to get the best of both worlds, but I'll close the issue as answered for now!