fix(anthropic): route OpenAI-compatible /v1/messages to chat/completions not Responses API

OpenAI-compatible deployments (custom_llm_provider=openai with a custom api_base
such as Zhipu / GLM) generally expose only chat/completions, so routing
/v1/messages to the Responses API since v1.92.0 produced 404 /responses errors.
Only genuine OpenAI endpoints now route to the Responses API; custom api_base
deployments fall back to chat/completions.

Fixes #33824

litellm/llms/anthropic/experimental_pass_through/messages/handler.py

@@ -19,6 +19,7 @@ from typing import (
     Union,
     cast,
 )
+from urllib.parse import urlparse
 
 import litellm
 from litellm.litellm_core_utils.litellm_logging import Logging as LiteLLMLoggingObj
@@ -31,6 +32,7 @@ from litellm.llms.base_llm.anthropic_messages.transformation import (
 )
 from litellm.llms.custom_httpx.http_handler import AsyncHTTPHandler
 from litellm.llms.custom_httpx.llm_http_handler import BaseLLMHTTPHandler
+from litellm.secret_managers.main import get_secret_str
 from litellm.types.llms.anthropic_messages.anthropic_request import AnthropicMetadata
 from litellm.types.llms.anthropic_messages.anthropic_response import (
     AnthropicMessagesResponse,
@@ -51,15 +53,36 @@ from .utils import AnthropicMessagesRequestUtils, mock_response
 _RESPONSES_API_PROVIDERS = frozenset({"openai"})
 
 
-def _should_route_to_responses_api(custom_llm_provider: Optional[str]) -> bool:
+def _is_openai_native_api_base(api_base: str | None) -> bool:
+    """Whether the effective OpenAI base URL points at the genuine OpenAI API.
+
+    OpenAI-compatible third parties (a custom ``api_base``, e.g. Zhipu / GLM)
+    generally expose only chat/completions, not the Responses API, so routing
+    /v1/messages there would hit a non-existent ``/responses`` path and 404.
+    Genuine OpenAI (no override, or an ``api.openai.com`` host) supports it.
+    """
+    effective = api_base or get_secret_str("OPENAI_API_BASE") or litellm.api_base
+    if not effective:
+        return True
+    hostname = urlparse(effective).hostname or ""
+    return hostname == "api.openai.com" or hostname.endswith(".api.openai.com")
+
+
+def _should_route_to_responses_api(custom_llm_provider: str | None, api_base: str | None) -> bool:
     """Return True when the provider should use the Responses API path.
 
     Set ``litellm.use_chat_completions_url_for_anthropic_messages = True`` to
     opt out and route OpenAI/Azure requests through chat/completions instead.
+
+    Only genuine OpenAI endpoints route to the Responses API. OpenAI-compatible
+    providers configured with a custom ``api_base`` fall back to chat/completions
+    since they typically do not implement ``/responses``.
     """
     if litellm.use_chat_completions_url_for_anthropic_messages:
         return False
-    return custom_llm_provider in _RESPONSES_API_PROVIDERS
+    if custom_llm_provider not in _RESPONSES_API_PROVIDERS:
+        return False
+    return _is_openai_native_api_base(api_base)
 
 
 def _deployment_passes_through_anthropic_messages(model_info: object) -> bool:
@@ -523,7 +546,7 @@ def anthropic_messages_handler(
             custom_llm_provider=custom_llm_provider,
             **kwargs,
         )
-        if _should_route_to_responses_api(custom_llm_provider):
+        if _should_route_to_responses_api(custom_llm_provider, api_base or dynamic_api_base):
             return LiteLLMMessagesToResponsesAPIHandler.anthropic_messages_handler(**_shared_kwargs)
 
         # The in-gateway context_management polyfill runs inside

tests/test_litellm/llms/anthropic/experimental_pass_through/messages/test_anthropic_experimental_pass_through_messages_handler.py

@@ -9,6 +9,7 @@ sys.path.insert(0, os.path.abspath("../../../../.."))
 
 from unittest.mock import AsyncMock, MagicMock, patch
 
+import litellm
 from litellm.anthropic_interface import messages
 from litellm.llms.custom_httpx.http_handler import AsyncHTTPHandler
 from litellm.types.utils import Delta, ModelResponse, StreamingChoices
@@ -821,3 +822,96 @@ def test_gate_passthrough_skipped_when_only_chat_completions_supported(monkeypat
     assert result == "translated"
     assert translation_calls["count"] == 1
     assert "config" not in captured
+
+
+def _responses_vs_completions_stubs(monkeypatch):
+    """Patch the two Anthropic->OpenAI translation handlers with distinct
+    return values so the caller can assert which endpoint a request routed to.
+    """
+    from litellm.llms.anthropic.experimental_pass_through.messages import handler
+
+    monkeypatch.setattr(
+        handler.LiteLLMMessagesToResponsesAPIHandler,
+        "anthropic_messages_handler",
+        staticmethod(lambda **kwargs: "responses"),
+    )
+    monkeypatch.setattr(
+        handler.LiteLLMMessagesToCompletionTransformationHandler,
+        "anthropic_messages_handler",
+        staticmethod(lambda **kwargs: "chat-completions"),
+    )
+
+
+@pytest.mark.parametrize(
+    "api_base",
+    [
+        "https://open.bigmodel.cn/api/paas/v4",
+        "https://api.groq.com/openai/v1",
+        "http://0.0.0.0:4000/v1",
+    ],
+)
+def test_openai_compatible_custom_api_base_routes_to_chat_completions(monkeypatch, api_base):
+    """Regression for #33824: an OpenAI-compatible deployment (provider=openai)
+    with a non-openai api_base must route /v1/messages through chat/completions,
+    not the Responses API which those endpoints do not implement (404 /responses).
+    """
+    from litellm.llms.anthropic.experimental_pass_through.messages.handler import (
+        anthropic_messages_handler,
+    )
+
+    _responses_vs_completions_stubs(monkeypatch)
+
+    result = anthropic_messages_handler(
+        max_tokens=100,
+        messages=[{"role": "user", "content": "Hello"}],
+        model="openai/glm-4.6",
+        api_key="sk-test",
+        api_base=api_base,
+    )
+
+    assert result == "chat-completions"
+
+
+@pytest.mark.parametrize("api_base", [None, "https://api.openai.com/v1"])
+def test_genuine_openai_routes_to_responses_api(monkeypatch, api_base):
+    """Genuine OpenAI endpoints (no override, or an api.openai.com host) still
+    route to the Responses API."""
+    monkeypatch.delenv("OPENAI_API_BASE", raising=False)
+    monkeypatch.setattr(litellm, "api_base", None)
+
+    from litellm.llms.anthropic.experimental_pass_through.messages.handler import (
+        anthropic_messages_handler,
+    )
+
+    _responses_vs_completions_stubs(monkeypatch)
+
+    result = anthropic_messages_handler(
+        max_tokens=100,
+        messages=[{"role": "user", "content": "Hello"}],
+        model="openai/gpt-5.1",
+        api_key="sk-test",
+        api_base=api_base,
+    )
+
+    assert result == "responses"
+
+
+def test_should_route_to_responses_api_honors_api_base():
+    from litellm.llms.anthropic.experimental_pass_through.messages.handler import (
+        _should_route_to_responses_api,
+    )
+
+    assert _should_route_to_responses_api("openai", None) is True
+    assert _should_route_to_responses_api("openai", "https://api.openai.com/v1") is True
+    assert _should_route_to_responses_api("openai", "https://us.api.openai.com/v1") is True
+    assert _should_route_to_responses_api("openai", "https://open.bigmodel.cn/api/paas/v4") is False
+    assert _should_route_to_responses_api("anthropic", None) is False
+
+
+def test_should_route_to_responses_api_respects_global_optout(monkeypatch):
+    from litellm.llms.anthropic.experimental_pass_through.messages.handler import (
+        _should_route_to_responses_api,
+    )
+
+    monkeypatch.setattr(litellm, "use_chat_completions_url_for_anthropic_messages", True)
+    assert _should_route_to_responses_api("openai", None) is False