fix(azure): rename max_tokens to max_completion_tokens for gpt-5-chat deployments (#36857)

Azure rejects the legacy `max_tokens` key for the whole gpt-5 name family, but
`AzureOpenAIGPT5Config.is_model_gpt_5_model` deliberately excludes `gpt-5-chat*`
so those deployments fall through to `AzureOpenAIConfig`, which sends `max_tokens`
verbatim and gets a 400 back on every request that carries it, `/health` probes
included.

One predicate was answering two independent questions. Split it: the new
`AzureOpenAIConfig.requires_max_completion_tokens` covers the whole gpt-5 name
family and drives only the rename, while `is_model_gpt_5_model` keeps keying
reasoning_effort, the temperature clamp and the dropped penalties off the
reasoning question, so #13781 stays fixed.

litellm/llms/azure/chat/gpt_transformation.py

@@ -112,6 +112,16 @@ class AzureOpenAIConfig(BaseConfig):
             "store",
         ]
 
+    @classmethod
+    def requires_max_completion_tokens(cls, model: str) -> bool:
+        """Whether Azure rejects the legacy ``max_tokens`` key for this deployment.
+
+        Deliberately wider than ``AzureOpenAIGPT5Config.is_model_gpt_5_model``: the whole gpt-5
+        name family needs the rename, including the ``gpt-5-chat*`` models that are excluded from
+        the reasoning path by https://github.com/BerriAI/litellm/issues/13781.
+        """
+        return "gpt-5" in model or "gpt5_series" in model
+
     def _is_response_format_supported_model(self, model: str) -> bool:
         """
         Determines if the model supports response_format.
@@ -160,6 +170,7 @@ class AzureOpenAIConfig(BaseConfig):
         api_version: str = "",
     ) -> dict:
         supported_openai_params: Final = self.get_supported_openai_params(model)
+        renames_max_tokens: Final = self.requires_max_completion_tokens(model)
         api_version_times: Final = api_version.split("-")
 
         if len(api_version_times) >= 3:
@@ -172,7 +183,9 @@ class AzureOpenAIConfig(BaseConfig):
             api_version_day = None
 
         for param, value in non_default_params.items():
-            if param == "tool_choice":
+            if param == "max_tokens" and renames_max_tokens:
+                optional_params.setdefault("max_completion_tokens", value)
+            elif param == "tool_choice":
                 """
                 This parameter requires API version 2023-12-01-preview or later
 

tests/test_litellm/llms/azure/chat/test_azure_chat_gpt_transformation.py

@@ -1,12 +1,21 @@
 import os
 import sys
+from typing import Final
+
+import pytest
+from pydantic import TypeAdapter
 
 sys.path.insert(
     0, os.path.abspath(os.path.join(os.path.dirname(__file__), "../../../../.."))
 )
 
+import litellm
 from litellm.litellm_core_utils.prompt_templates.common_utils import TOOL_RESULT_IMAGE_BOUNDARY
 from litellm.llms.azure.chat.gpt_transformation import AzureOpenAIConfig
+from litellm.utils import get_optional_params
+
+_MAPPED_PARAMS: Final = TypeAdapter(dict[str, object])
+_SUPPORTED_PARAMS: Final = TypeAdapter(list[str])
 
 
 class TestAzureOpenAIConfig:
@@ -91,3 +100,69 @@ def test_transform_request_hoists_tool_message_image():
         {"type": "text", "text": TOOL_RESULT_IMAGE_BOUNDARY},
         {"type": "image_url", "image_url": {"url": data_uri}},
     ]
+
+
+@pytest.mark.parametrize(
+    "model, emitted_key, absent_key",
+    [
+        ("gpt-5-chat", "max_completion_tokens", "max_tokens"),
+        ("gpt-5-chat-latest", "max_completion_tokens", "max_tokens"),
+        ("gpt-5-chat-2025-08-07", "max_completion_tokens", "max_tokens"),
+        ("gpt-5", "max_completion_tokens", "max_tokens"),
+        ("o3-mini", "max_completion_tokens", "max_tokens"),
+        ("gpt-4o", "max_tokens", "max_completion_tokens"),
+    ],
+)
+def test_azure_max_tokens_rename_covers_gpt_5_chat_family(model: str, emitted_key: str, absent_key: str) -> None:
+    """Azure rejects `max_tokens` for the whole gpt-5 name family, gpt-5-chat* included."""
+    mapped: Final = _MAPPED_PARAMS.validate_python(
+        get_optional_params(model=model, custom_llm_provider="azure", max_tokens=5)
+    )
+    assert mapped[emitted_key] == 5
+    assert absent_key not in mapped
+
+
+@pytest.mark.parametrize("model", ["gpt-5-chat", "gpt-5-chat-latest"])
+def test_azure_gpt_5_chat_stays_off_the_reasoning_path(model: str) -> None:
+    """https://github.com/BerriAI/litellm/issues/13781: gpt-5-chat* is a regular chat model."""
+    mapped: Final = _MAPPED_PARAMS.validate_python(
+        get_optional_params(
+            model=model,
+            custom_llm_provider="azure",
+            max_tokens=5,
+            temperature=0.3,
+            presence_penalty=0.1,
+            frequency_penalty=0.2,
+            stop=["stop"],
+            logit_bias={"1": 1},
+        )
+    )
+    supported: Final = _SUPPORTED_PARAMS.validate_python(
+        litellm.get_supported_openai_params(model=model, custom_llm_provider="azure")
+    )
+    assert mapped["temperature"] == 0.3
+    assert mapped["presence_penalty"] == 0.1
+    assert mapped["frequency_penalty"] == 0.2
+    assert mapped["stop"] == ["stop"]
+    assert mapped["logit_bias"] == {"1": 1}
+    assert "reasoning_effort" not in mapped
+    assert "reasoning_effort" not in supported
+
+
+def test_azure_gpt_5_takes_the_reasoning_path() -> None:
+    """Positive control for the predicate split: gpt-5 still drops chat-only params."""
+    mapped: Final = _MAPPED_PARAMS.validate_python(
+        get_optional_params(
+            model="gpt-5",
+            custom_llm_provider="azure",
+            presence_penalty=0.1,
+            logit_bias={"1": 1},
+            drop_params=True,
+        )
+    )
+    supported: Final = _SUPPORTED_PARAMS.validate_python(
+        litellm.get_supported_openai_params(model="gpt-5", custom_llm_provider="azure")
+    )
+    assert "presence_penalty" not in mapped
+    assert "logit_bias" not in mapped
+    assert "reasoning_effort" in supported