BaseModel output type no longer working for Gemini models

[pjsample]

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

I'm running into a new Gemini issue using BaseModel output types. Code that was working yesterday is now returning an error.

ValidationError: 1 validation error for _GeminiResponse
candidates.0.content.parts.0.text.text
  Field required [type=missing, input_value={'thought': True, 'though...Kme/E40jb0xxL+qYyMsw=='}, input_type=dict]

OpenAI models are fine so it's Gemini specific.

Working:

agent = Agent(
    model="google-gla:gemini-2.5-flash-preview-05-20",
    output_type=str
)

result = await agent.run("What's up?")
print(result.output)

Not working:

class SimpleResponse(BaseModel):
    output: str = Field(description="The output of the LLM.")

agent = Agent(
    model="google-gla:gemini-2.5-flash-preview-05-20",
    output_type=SimpleResponse
)

result = await agent.run("What's up?")
print(result.output)

Example Code

class SimpleResponse(BaseModel):
    output: str = Field(description="The output of the LLM.")

agent = Agent(
    model="google-gla:gemini-2.5-flash-preview-05-20",
    output_type=SimpleResponse
)

result = await agent.run("What's up?")
print(result.output)

Python, Pydantic AI & LLM client version

Python 3.11.11
pydantic-ai-slim                         0.2.12
pydantic-core                            2.33.1
pydantic-graph                           0.2.12
pydantic                                 2.11.3

[MacroMackie]

This is affecting us too - seems to specifically affect us in situations where: thinking_budget>0.

The gemini api seems to be returning:

parts: [
    {'thought': True, 'thoughtSignature': 'AZs1mZxelyMPKoHcvslvjy5WZYI2HosTK...'},
    {'functionCall': {'name': 'final_result_XyzResponse', 'args': {'content': 'example'}}}
]

and that first part is causing it to raise the validation error.

[DouweM]

@MacroMackie That's right.

As a temporary workaround, you can disable thinking: https://ai.pydantic.dev/models/gemini/#disable-thinking

@Kludex Can you please have a look at this as you've been working on Google and Thinking?

[omarbaker8]

switching to GoogleModel class rather than GeminiModel would solve it

[nguyen-tran-ds-vns]

switching to GoogleModel class rather than GeminiModel would solve it

Can you show us how? I can't seem to find the docs showing how to use GoogleModel class.

[nguyen-tran-ds-vns]

@MacroMackie That's right.

As a temporary workaround, you can disable thinking: https://ai.pydantic.dev/models/gemini/#disable-thinking

@Kludex Can you please have a look at this as you've been working on Google and Thinking?

Does work for 'gemini-2.5-flash-preview-04-17'
Still validation error for 'gemini-2.5-pro-preview-05-06'

[akhilbatra898]

from pydantic_ai.models.google import GoogleModel
model = GoogleModel("gemini-2.5-flash-preview-04-17")

agent = Agent(model)

Snippet to use Google Model

[amiyapatanaik]

It appears Google suddenly decided switch on thoughts in API (thought summaries to be precise). More detail here: https://discuss.ai.google.dev/t/massive-regression-detailed-gemini-thinking-process-vanished-from-ai-studio/83916/103
Disabling thinking can adversely affect the outputs, so for now using Google Model is a good work around.

[nguyen-tran-ds-vns]

@akhilbatra898 Can help point out what I get wrong here?

gemini_2_5_pro = GoogleModel('gemini-2.5-pro-preview-05-06',
    provider=GoogleGLAProvider(api_key=settings.gemini_key)
)

Agent(model=gemini_2_5_pro).run_sync('hello')

[/usr/local/lib/python3.11/dist-packages/pydantic_ai/models/google.py](https://localhost:8080/#) in _generate_content(self, messages, stream, model_settings, model_request_parameters)
    270         )
    271 
--> 272         func = self.client.aio.models.generate_content_stream if stream else self.client.aio.models.generate_content
    273         return await func(model=self._model_name, contents=contents, config=config)  # type: ignore
    274 

AttributeError: 'AsyncClient' object has no attribute 'aio'

[ahstn]

@nguyen-tran-ds-vns - yeah seeing the same when trying GoogleModel:

model = GoogleModel(
    model_id, # resolves to "gemini-2.5-flash-preview-05-20"
    provider=GoogleVertexProvider(service_account_file=GCP_CREDENTIALS_PATH),
)
agent = Agent(
    model,
    output_type=list[MyModel],
    output_retries=3,

    # Can't disable thinking for 2.5 Flash
    # model_settings=GeminiModelSettings(gemini_thinking_config={"thinking_budget": 0}),,
)
result = await agent.run(prompt, output_type=list[MyModel])
return result

Stack Trace:

AttributeError: 'AsyncClient' object has no attribute 'aio'
  File "/app/gcp_vertex.py", line 45, in assign_product_category
    result = await agent.run(prompt, output_type=list[MyModel])
             ^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^
  File "/app/.venv/lib/python3.11/site-packages/pydantic_ai/agent.py", line 468, in run
    async for _ in agent_run:
  File "/app/.venv/lib/python3.11/site-packages/pydantic_ai/agent.py", line 1923, in __anext__
    next_node = await self._graph_run.__anext__()
                ^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^
  File "/app/.venv/lib/python3.11/site-packages/pydantic_graph/graph.py", line 809, in __anext__
    return await self.next(self._next_node)
           ^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^
  File "/app/.venv/lib/python3.11/site-packages/pydantic_graph/graph.py", line 782, in next
    self._next_node = await node.run(ctx)
                      ^^^^^^^^^^^^^^^^^^^
  File "/app/.venv/lib/python3.11/site-packages/pydantic_ai/_agent_graph.py", line 292, in run
    return await self._make_request(ctx)
           ^^^^^^^^^^^^^^^^^^^^^^^^^^^^^
  File "/app/.venv/lib/python3.11/site-packages/pydantic_ai/_agent_graph.py", line 345, in _make_request
    model_response = await ctx.deps.model.request(
                     ^^^^^^^^^^^^^^^^^^^^^^^^^^^^^
  File "/app/.venv/lib/python3.11/site-packages/pydantic_ai/models/google.py", line 174, in request
    response = await self._generate_content(messages, False, model_settings, model_request_parameters)
               ^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^
  File "/app/.venv/lib/python3.11/site-packages/pydantic_ai/models/google.py", line 280, in _generate_content
    func = self.client.aio.models.generate_content_stream if stream else self.client.aio.models.generate_content
                                                                         ^^^^^^^^^^^^^^^

Which I believe is this PR #1751 - pydantic_ai_slim/pydantic_ai/models/google.py

[strunov]

Find it strange that setting GeminiModelSettings(gemini_thinking_config=ThinkingConfig(include_thoughts=False)) also doesn't work - thoughts are still returned by the vertex model

So far setting budget to 0 is the only thing that worked, but it regresses the response