fix(azure_ai): redirect a gpt-5 capability lookup only when the map has a foundry row

gpt-6-astra is the only gpt-5-family name with an azure_ai row. Prefixing the rest
cost them every effort flag, since get_llm_provider sends an azure_ai name down the
azure provider when a global AZURE_AI_API_BASE points at an openai.azure.com host and
azure/<model> is not a key either, which turned temperature, top_p and logprobs on
azure_ai/gpt-5.1-chat-latest from accepted into an UnsupportedParamsError.

litellm/llms/azure_ai/chat/transformation.py

@@ -46,7 +46,19 @@ NON_OPENAI_SPEC_MESSAGE_FIELDS: Final = (
 class AzureAIGPT5Config(OpenAIGPT5Config):
     @classmethod
     def _model_map_lookup_name(cls, model: str) -> str:
-        return model if model.startswith("azure_ai/") else f"azure_ai/{model}"
+        """Normalise a Foundry routing name to its cost-map key, when the map has one.
+
+        A Foundry deployment and its OpenAI-hosted namesake are different products with
+        different capabilities, so ``azure_ai/<model>`` is the entry to read whenever the map
+        carries it. Most gpt-5-family names have no ``azure_ai/`` row, though, and prefixing
+        those anyway costs them every flag: ``get_llm_provider`` re-resolves an ``azure_ai/``
+        name to the azure provider when a global AZURE_AI_API_BASE points at an
+        openai.azure.com host, ``azure/<model>`` is not a key either, so the lookup lands
+        nowhere and every effort answer degrades to False. A missing key defers to the base
+        resolver instead.
+        """
+        prefixed: Final = model if model.startswith("azure_ai/") else f"azure_ai/{model}"
+        return prefixed if prefixed in litellm.model_cost else super()._model_map_lookup_name(model)
 
 
 azureAIGPT5Config: Final = AzureAIGPT5Config()

tests/test_litellm/llms/azure_ai/chat/test_azure_ai_transformation.py

@@ -157,6 +157,29 @@ def test_foundry_gpt_6_astra_keeps_sampling_params_when_reasoning_effort_is_none
     assert optional_params == {"reasoning_effort": "none", "temperature": 0.2, "top_p": 0.9}
 
 
+def test_a_gpt_5_name_without_a_foundry_row_keeps_reading_its_own_entry(
+    monkeypatch: pytest.MonkeyPatch, _local_model_cost_map
+):
+    """gpt-6-astra is the only gpt-5-family name with an azure_ai/ row. Reading an azure_ai/ key for
+    the rest finds nothing, and an openai.azure.com base sends that name down the azure provider,
+    which has no key for it either, so every effort answer would silently fall back to false and
+    take temperature, top_p and logprobs down with it."""
+    monkeypatch.setenv("AZURE_AI_API_BASE", "https://example-resource.openai.azure.com")
+    monkeypatch.setenv("AZURE_AI_API_KEY", "placeholder")
+
+    optional_params = litellm.utils.get_optional_params(
+        model="gpt-5.1-chat-latest",
+        custom_llm_provider="azure_ai",
+        temperature=0.2,
+        top_p=0.9,
+        logprobs=True,
+    )
+
+    assert optional_params["temperature"] == 0.2
+    assert optional_params["top_p"] == 0.9
+    assert optional_params["logprobs"] is True
+
+
 def test_azure_ai_grok_stop_parameter_handling():
     """
     Test that Grok models properly handle stop parameter filtering in Azure AI Studio.