thinking_level not accepted when initializing Gemini 3 via init_chat_model()

[Gu-Haojia]

Package (Required)

• langchain-google-genai
• langchain-google-vertexai
• langchain-google-community
• Other / not sure / general

Checked other resources

• I added a descriptive title to this issue
• I searched the LangChain documentation and API reference (linked above)
• I used the GitHub search to find a similar issue and didn't find it
• I am sure this is a bug and not a question or request for help

Example Code (Python)

from langchain.chat_models import init_chat_model

model = init_chat_model(
    "google_genai:gemini-3-pro-preview",
    thinking_level="low",  # <-- new field from Gemini docs
)

# do invoke

Error Message and Stack Trace (if applicable)

Model return: Unknown field for ThinkingConfig: thinking_level

Description

According to Google’s latest Gemini thinking docs (https://ai.google.dev/gemini-api/docs/thinking), thinking_level is a supported parameter alongside thinking_budget. When we attempt to pass thinking_level="low" through LangChain’s init_chat_model, the config validation rejects it with “Unknown field for ThinkingConfig: thinking_level”. This prevents us from using the new low/high thinking presets while still leveraging LangChain’s tooling/graph integrations.

Environment: langchain==1.0.7, langchain-core==1.0.5, langchain-google-genai==3.1.0.
Expectation: LangChain should either accept thinking_level and forward it to Gemini, or clearly document that only the legacy thinking_budget* fields are supported.

[Yuki-Imajuku]

Same here.
This is because langchain-gogle-genai uses google-ai-generativelanguage, not google-genai

[FedericoCampe8]

cc @ccurme possibly related to this? #1365

[Yuki-Imajuku]

I noticed someone address migration to google-genai. #1330

[mbn18]

I tested it:

llm = ChatGoogleGenerativeAI(model='gemini-3-pro', media_resolution=3, thinking_level="low")
config = llm._prepare_params(stop=None)

Receiving the following error:

ValueError: Unknown field for ThinkingConfig: thinking_level

I took the code from the test file