fix(openai): Fix Azure OpenAI Responses API model field issue (#32649)

libs/partners/openai/langchain_openai/chat_models/azure.py

@@ -754,6 +754,26 @@ class AzureChatOpenAI(BaseChatOpenAI):
 
         return chat_result
 
+    def _get_request_payload(
+        self,
+        input_: LanguageModelInput,
+        *,
+        stop: Optional[list[str]] = None,
+        **kwargs: Any,
+    ) -> dict:
+        """Get the request payload, using deployment name for Azure Responses API."""
+        payload = super()._get_request_payload(input_, stop=stop, **kwargs)
+
+        # For Azure Responses API, use deployment name instead of model name
+        if (
+            self._use_responses_api(payload)
+            and not payload.get("model")
+            and self.deployment_name
+        ):
+            payload["model"] = self.deployment_name
+
+        return payload
+
     def _stream(self, *args: Any, **kwargs: Any) -> Iterator[ChatGenerationChunk]:
         """Route to Chat Completions or Responses API."""
         if self._use_responses_api({**kwargs, **self.model_kwargs}):

libs/partners/openai/langchain_openai/chat_models/base.py

@@ -3541,7 +3541,7 @@ def _construct_responses_api_payload(
         payload["reasoning"] = {"effort": payload.pop("reasoning_effort")}
 
     # Remove temperature parameter for models that don't support it in responses API
-    model = payload.get("model", "")
+    model = payload.get("model") or ""
     if model.startswith("gpt-5") and "chat" not in model:  # gpt-5-chat supports
         payload.pop("temperature", None)
 

libs/partners/openai/tests/integration_tests/chat_models/test_azure_standard.py

@@ -21,7 +21,6 @@ class TestAzureOpenAIStandard(ChatModelIntegrationTests):
     def chat_model_params(self) -> dict:
         return {
             "deployment_name": os.environ["AZURE_OPENAI_CHAT_DEPLOYMENT_NAME"],
-            "model": "gpt-4o-mini",
             "openai_api_version": OPENAI_API_VERSION,
             "azure_endpoint": OPENAI_API_BASE,
             "stream_usage": True,
@@ -49,7 +48,6 @@ class TestAzureOpenAIResponses(ChatModelIntegrationTests):
     def chat_model_params(self) -> dict:
         return {
             "deployment_name": os.environ["AZURE_OPENAI_CHAT_DEPLOYMENT_NAME"],
-            "model": "gpt-4o-mini",
             "openai_api_version": OPENAI_API_VERSION,
             "azure_endpoint": OPENAI_API_BASE,
             "use_responses_api": True,

libs/partners/openai/tests/unit_tests/chat_models/test_azure.py

@@ -5,6 +5,7 @@ from unittest import mock
 
 import pytest
 from langchain_core.messages import HumanMessage
+from pydantic import SecretStr
 from typing_extensions import TypedDict
 
 from langchain_openai import AzureChatOpenAI
@@ -99,3 +100,42 @@ def test_max_completion_tokens_in_payload() -> None:
         "stream": False,
         "max_completion_tokens": 300,
     }
+
+
+def test_responses_api_uses_deployment_name() -> None:
+    """Test that Azure deployment name is used for Responses API."""
+    llm = AzureChatOpenAI(
+        azure_deployment="your_deployment",
+        api_version="2025-04-01-preview",
+        azure_endpoint="your_endpoint",
+        api_key=SecretStr("your_api_key"),
+        # Force Responses API usage by including a Responses-only parameter
+        use_responses_api=True,
+        output_version="responses/v1",
+    )
+    messages = [HumanMessage("Hello")]
+    payload = llm._get_request_payload(messages)
+
+    # For Responses API, the model field should be the deployment name
+    assert payload["model"] == "your_deployment"
+    assert "input" in payload  # Responses API uses 'input' instead of 'messages'
+
+
+def test_chat_completions_api_uses_model_name() -> None:
+    """Test that regular Chat Completions API still uses model name."""
+    llm = AzureChatOpenAI(
+        azure_deployment="your_deployment",
+        model="gpt-5",  # This is the OpenAI model name
+        api_version="2025-04-01-preview",
+        azure_endpoint="your_endpoint",
+        api_key=SecretStr("your_api_key"),
+        # No Responses-only parameters, so Chat Completions API will be used
+    )
+    messages = [HumanMessage("Hello")]
+    payload = llm._get_request_payload(messages)
+
+    # For Chat Completions API, the model field should still be None/model_name
+    # Azure Chat Completions uses deployment in the URL, not in the model field
+    assert payload["model"] == "gpt-5"
+    assert "messages" in payload  # Chat Completions API uses 'messages'
+    assert "input" not in payload