Getting "Received empty model response" when using GPT 5 model series

[praveen92y]

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 GPT 5 series model getting below exception:

venv/lib/python3.13/site-packages/pydantic_ai/_agent_graph.py", line 488, in _run_stream
     async for event in self._events_iterator:
         yield event
venv/lib/python3.13/site-packages/pydantic_ai/_agent_graph.py", line 484, in _run_stream
     raise exceptions.UnexpectedModelBehavior('Received empty model response')
 pydantic_ai.exceptions.UnexpectedModelBehavior: Received empty model response

While searching over web, i have found below solution:

where they mentioned:

from openai import AsyncOpenAI

client = AsyncOpenAI()

response = await client.responses.create(
    model="gpt-5-mini",
    input=[
        {"role": "system", "content": "You are a helpful assistant."},
        {"role": "user", "content": "Say hello in one sentence."}
    ],
    max_output_tokens=50,
)

print(f"output_text: '{response.output_text}'")  # Empty!
print(f"output[0] type: {response.output[0].type}")  # reasoning

Temporary Workaround
Force the output format to text so the model returns a message item:

response = await client.responses.create(
    model="gpt-5-mini",
    input=[
        {"role": "system", "content": "You are a helpful assistant."},
        {"role": "user", "content": "Say hello in one sentence."}
    ],
    text={"format": {"type": "text"}},
    max_output_tokens=200,
)

print(response.output_text)  # Works now

https://community.openai.com/t/gpt-5-mini-responses-api-empty-output-text-issue/1342786

So, i think this needs to be handled in pydanticAI as well for GPT series model.

Example Code

Python, Pydantic AI & LLM client version

Python version: 3.11
Pydantic: 0.7.4
LLM Client : OpenAIModel(
                'gpt-5-mini',
                provider=AzureProvider(
                    azure_endpoint=azure_endpoint,
                    api_version=api_version,
                    api_key=api_key,
                ),
            )

[pydanticai-bot]

PydanticAI Github Bot Found 1 issues similar to this one:

1. "Getting "Received empty model response" when using GPT 5 model series #2644" (95% similar)

[praveen92y]

PydanticAI Github Bot Found 1 issues similar to this one:

1. "Getting "Received empty model response" when using GPT 5 model series #2644" (95% similar)

That was actually created by me, but i put that on Closed as not planned because that was not having enough info at that time

[DouweM]

@praveen92y Thanks for the report! Can you share minimal Pydantic AI that reproduces this issue as well, please?

It sounds like we should always send text={"format": {"type": "text"}} to the Responses API in an else branch after these branches for JSON format:

pydantic-ai/pydantic_ai_slim/pydantic_ai/models/openai.py

Lines 874 to 889 in 057c0bc

	text: responses.ResponseTextConfigParam | None = None
	if model_request_parameters.output_mode == 'native':
	output_object = model_request_parameters.output_object
	assert output_object is not None
	text = {'format': self._map_json_schema(output_object)}
	elif (
	model_request_parameters.output_mode == 'prompted' and self.profile.supports_json_object_output
	): # pragma: no branch
	text = {'format': {'type': 'json_object'}}
	# Without this trick, we'd hit this error:
	# > Response input messages must contain the word 'json' in some form to use 'text.format' of type 'json_object'.
	# Apparently they're only checking input messages for "JSON", not instructions.
	assert isinstance(instructions, str)
	openai_messages.insert(0, responses.EasyInputMessageParam(role='system', content=instructions))
	instructions = NOT_GIVEN

We'd need to verify that

• that doesn't conflict with tool_choice=required (as in that case there can be no text output anyway)
• that this doesn't cause issues with any models prior to GPT-5.

If it does, we'll need some additional checks.

Would you be up for submitting a PR to fix this?

[g-eoj]

I can repro the error reliably by setting max_tokens=10 or some low value. I don't seem to get the error with higher max_token values. Whatever the default is it would give me an empty response.

@praveen92y, if you set a high max_tokens does the error go away?

Edit: for reference vllm-project/vllm#24184

[DouweM]

@g-eoj Yeah the issue seems to be just be that the max tokens is too low for the model to actually be able to fulfill the prompt, and {"format": {"type": "text"}} doesn't seem to make a difference.

I'll close this issue unless/until we can reproduce this with a valid example without an arbitrarily low max tokens.

[g-eoj]

@DouweM I was just writing a new comment: I think both fixes are required for vLLM. I would test the OpenAI endpoints but I don't want to pay for an account.

I still eventually hit the empty model response error unless I patch Pydantic AI as described here: #2645 (comment)

[DouweM]

@g-eoj Hmm, {"format": {"type": "text"}} fixes it for vLLM? Can you please create a new issue for that? This issue is not really actionable as the workaround in the description really just works because of max_output_tokens=200, not text={"format": {"type": "text"}}.

[g-eoj]

@DouweM I'm not sure if the fix is actually a fix or just pushes the error somewhere else. Now I'm finding other errors. I'm going to keep looking into it, will make an issue if I can come up with an actionable repro of an error.