fix(responses): promote OpenAI-SDK base_url to internal api_base

Fixes #30026.

litellm.completion() converts the OpenAI-SDK-style 'base_url' kwarg into
the internal 'api_base' before building LitellmParams (main.py:1384).
litellm.responses() didn't, so YAML model_list entries that set
base_url: https://eu.api.openai.com/v1 silently hit api.openai.com on
the /v1/responses path while /v1/chat/completions on the same deployment
worked fine.

Add the same one-line promotion in responses() just before constructing
GenericLiteLLMParams. api_base passed explicitly still wins; absence of
both leaves the kwargs untouched.

litellm/responses/main.py

@@ -954,6 +954,14 @@ def responses(
             # Update local_vars to include the converted text parameter
             local_vars["text"] = text
 
+        # Mirror completion()'s OpenAI-SDK-style base_url -> internal api_base promotion
+        # (litellm/main.py:1384). YAML model_list / OpenAI client kwargs use 'base_url';
+        # CredentialLiteLLMParams stores 'api_base'. Without this, GenericLiteLLMParams
+        # below leaves api_base=None and the request goes to api.openai.com regardless of
+        # the configured base_url (issue #30026).
+        if kwargs.get("base_url") is not None and kwargs.get("api_base") is None:
+            kwargs["api_base"] = kwargs.pop("base_url")
+
         # get llm provider logic
         litellm_params = GenericLiteLLMParams(**kwargs)
 

tests/test_litellm/responses/test_responses_base_url_passthrough.py

@@ -0,0 +1,68 @@
+"""Regression for #30026 — proxy YAML's ``base_url`` (OpenAI-SDK convention) must reach
+the OpenAI client as ``api_base``. Before the fix, ``litellm.responses()`` constructed
+``GenericLiteLLMParams(**kwargs)`` without first mapping ``base_url`` to ``api_base``,
+so requests went to ``api.openai.com`` regardless of the configured base_url."""
+
+import importlib
+from unittest.mock import patch
+
+import pytest
+
+import litellm
+from litellm.types.router import GenericLiteLLMParams
+
+# Import the responses package's main module directly; `litellm.responses` is a
+# callable in the public namespace, so we can't dotted-import through it.
+responses_main = importlib.import_module("litellm.responses.main")
+
+
+def _patched_params_capture(monkeypatch):
+    """Return a (seen-dict, patched-class) pair that records the kwargs handed to
+    GenericLiteLLMParams inside responses()."""
+    seen = {}
+
+    class _Spy(GenericLiteLLMParams):
+        def __init__(self, **data):
+            seen.update(data)
+            super().__init__(**data)
+
+    monkeypatch.setattr(responses_main, "GenericLiteLLMParams", _Spy)
+    return seen
+
+
+def test_base_url_in_kwargs_is_promoted_to_api_base(monkeypatch):
+    seen = _patched_params_capture(monkeypatch)
+    # mock_response short-circuits in responses() right after litellm_params is built,
+    # so we reach the promotion logic without making an HTTP call.
+    litellm.responses(
+        input="hello",
+        model="openai/gpt-4o-mini",
+        base_url="https://eu.api.openai.com/v1",
+        mock_response="ok",
+    )
+    assert seen.get("api_base") == "https://eu.api.openai.com/v1"
+    assert "base_url" not in seen
+
+
+def test_existing_api_base_takes_precedence(monkeypatch):
+    seen = _patched_params_capture(monkeypatch)
+    litellm.responses(
+        input="hi",
+        model="openai/gpt-4o-mini",
+        base_url="https://wrong.example.com/v1",
+        api_base="https://right.example.com/v1",
+        mock_response="ok",
+    )
+    assert seen.get("api_base") == "https://right.example.com/v1"
+
+
+def test_no_base_url_leaves_api_base_unset(monkeypatch):
+    seen = _patched_params_capture(monkeypatch)
+    litellm.responses(
+        input="hi",
+        model="openai/gpt-4o-mini",
+        mock_response="ok",
+    )
+    # Neither key present → no promotion, api_base stays absent from the kwargs.
+    assert "api_base" not in seen
+    assert "base_url" not in seen