diff --git a/strix/config/models.py b/strix/config/models.py
index dfd54cb9..84a42deb 100644
--- a/strix/config/models.py
+++ b/strix/config/models.py
@@ -736,10 +736,13 @@ def resolve_api_type(model_name: str, settings: Settings) -> ApiType:
 
     An explicit ``STRIX_API_TYPE`` wins. Otherwise the model decides: Responses
     when LiteLLM's catalog lists ``/v1/responses`` for it, chat completions for
-    everything else.
+    everything else. Models routed through LiteLLM (non-OpenAI providers) always
+    use chat completions because the Responses endpoint is OpenAI-specific.
     """
     if settings.llm.api_type is not None:
         return settings.llm.api_type
+    if routes_through_litellm(model_name):
+        return "chat_completions"
     if _RESPONSES_ENDPOINT in _catalog_supported_endpoints(model_name):
         return "responses"
     return "chat_completions"
diff --git a/tests/test_models.py b/tests/test_models.py
index aafbaf40..fc2f8f31 100644
--- a/tests/test_models.py
+++ b/tests/test_models.py
@@ -169,6 +169,14 @@ def test_resolve_api_type_follows_the_catalog_not_the_base_url(
         assert uses_chat_completions_tool_schema(model, settings) is True, model
 
 
+def test_litellm_models_use_chat_completions_even_when_catalog_lists_responses(
+    monkeypatch: pytest.MonkeyPatch,
+) -> None:
+    monkeypatch.setattr(models, "_catalog_supported_endpoints", lambda _model: ["/v1/responses"])
+
+    assert resolve_api_type("anthropic/claude-sonnet-4-5", Settings()) == "chat_completions"
+
+
 def test_resolve_api_type_explicit_override_wins(monkeypatch: pytest.MonkeyPatch) -> None:
     _settings(monkeypatch, "gpt-daybreak-blue-latest", "https://gateway.example/v1")
     monkeypatch.setenv("STRIX_API_TYPE", "chat_completions")