LangChain Python

ChatOpenAI from langchain-openai. For Claude through the Messages API use LangChain Python (ChatAnthropic); for Gemini, LangChain Google GenAI.

Connect

def chat_model(model, **overrides):
    from langchain_openai import ChatOpenAI

    settings = {
        "base_url": os.environ["PLUGBOARD_BASE_URL"],  # ends in /v1
        "api_key": os.environ["PLUGBOARD_KEY"],
        "model": model,
        "use_responses_api": False,  # Chat Completions. Left unset, a gpt-5.6 id can switch to /v1/responses
        **NO_RETRIES,
    }
    if model == "fast":
        settings["max_tokens"] = 1024  # fast is a reasoning model: leave room for the answer after its thinking
    settings.update(overrides)
    return ChatOpenAI(**settings)

Set use_responses_api=False to stay on Chat Completions: otherwise some model ids (gpt-5.6...) switch the client to the Responses API.

Chat

message = client(model).invoke("Reply with exactly: pong")

Tool calls

@tool
def get_weather(city: str) -> str:
    """Current weather for a city."""
    return TOOL_RESULT

llm = client(model).bind_tools([get_weather])
first = llm.invoke(messages)
call = first.tool_calls[0]  # {"name", "args", "id"}; the id goes back on the ToolMessage
messages += [first, ToolMessage(content=get_weather.invoke(call["args"]), tool_call_id=call["id"])]
second = llm.invoke(messages)

Structured output

llm.with_structured_output(Meal, method="json_schema") and method="function_calling" both work. Avoid method="json_mode": it sends no schema, so the model guesses the keys.

Responses API

ChatOpenAI(use_responses_api=True, ...) sends the same calls to /v1/responses:

def responses_model(model, **overrides):
    """The same client on the Responses API (`/v1/responses`). Set explicitly: LangChain also switches on its own for gpt-5.6 ids."""
    return chat_model(model, use_responses_api=True, **overrides)

On that route LangChain exposes no finish_reason; read response_metadata["status"] instead.

Limits

• LangChain has no model-list call. Use the wrapped client: llm.root_client.models.list().
• Errors are the openai package's: AuthenticationError (401), NotFoundError (404).