Question on how to use Gemini 3 model through a custom gateway

[zuoxiang95]

I would like to use Gemini3 with a custom gateway address, but currently it only supports the /responses and does not support interfaces related to /chat/completions. May I ask how to configure the agent to use the /responses endpoint?

[zuoxiang95]

I have used set_default_openai_api("responses"), but it doesn't work.

custom_client = AsyncOpenAI(
    base_url="***",
    api_key="****"
)
set_default_openai_api("responses")
my_model = OpenAIChatCompletionsModel(
        model="Gemini 3-Pro-Preview",
        openai_client=custom_client,
)

[ihower]

@zuoxiang95 You can refer to my setup. There are two approaches.

1. Use Gemini’s OpenAI-compatible endpoint: https://github.com/ihower/openai-agents-python-live-tests/blob/main/chatcmpl-gemini3.py

2. Use LiteLLM: https://github.com/ihower/openai-agents-python-live-tests/blob/main/litellm-gemini3.py

That said, Gemini 3 in the Agents SDK is not fully support yet. Please see this PR: #2158

[seratch]

As @ihower mentioned, you can use only Chat Completions API interface for agents backed by Gemini 3 model.