Disable thinking mode in Gemini 2.5 flash

[tcrapts]

Description

Gemini 2.5 flash has thinking mode enabled by default. This is is expensive. It can be turned off, but Pydantic AI doesn't support this.

AISDK from Vercel does have this. We need something similar:

References

No response

[Kludex]

I'm working on this already. Should be ready at some point this week.

[bigs]

Hey @Kludex excited to see this coming! Was wondering if, while you were working on thinking options, if #840 might also be relevant as a model-specific option? Just throwing it into your consciousness!

[amiyapatanaik]

A small example of how to achieve this will be very helpful.

[tcrapts]

@amiyapatanaik refer to the docs; https://ai.pydantic.dev/api/models/gemini/#pydantic_ai.models.gemini.GeminiModelSettings

Something like this should work:

from pydantic_ai import Agent
from pydantic_ai.models.gemini import GeminiModel, ThinkingConfig
from pydantic_ai.providers.google_gla import GoogleGLAProvider

api_key = "your_key"

gemini_thinking_config: ThinkingConfig = {
    "include_thoughts": False,
    "thinking_budget": 0,
}


agent = Agent(
    model=GeminiModel(
        model_name="gemini-2.5-flash-preview-04-17",
        provider=GoogleGLAProvider(
            api_key=api_key,
        ),
    ),
    model_settings={"gemini_thinking_config": gemini_thinking_config},  # type: ignore
)

result = agent.run_sync("Who are you?")
print(result.output)

I need to ignore type issues, because the default ModelSettings class somehow doesn't seem to recognize gemini_thinking_config. But it works.

[d-zimmermann]

I expected to give do something along the line of:

# Disable thinking (budget 0) and output; only works for GEMINI
thinkingDisabledConfig = ThinkingConfig(
    include_thoughts=False,
    thinking_budget=0
)

model_settings=ModelSettings(    
    thinking_config=thinkingDisabledConfig,
    temperature=1.0, 
    top_p=0.95,
    top_k=64,
)

agent = Agent(  
            'google-gla:gemini-2.5-flash-preview-04-17',
            model_settings=model_settings
)

... but this is completely ignored and for Usage(), I can see that it's thinking along anyway.

The SOLUTION is using gemini_thinking_config in ModelSettings.

model_settings=ModelSettings(    
    gemini_thinking_config=thinkingDisabledConfig,
    temperature=1.0, 
    top_p=0.95,
    top_k=64,
)

Not sure if this is intended or a bug though. It makes it unnecessarily complicated to use a different model than Gemini with this behaviour. I had hoped that it would work with generic ModelSettings and model-specific settings, and that these would just be ignored if not supported. ('gemini' prefix)

EDIT: Thanks @tcrapts for pointing this out!

[tcrapts]

@d-zimmermann ModelSettings and GeminiModelSettings are both TypedDict, in runtime it doesn't matter which you use. I think both of your examples will not work, the correct key is gemini_thinking_config, with the gemini_ prefix.