fix: drop effort parameter for Haiku with azure_ai provider

litellm/llms/azure_ai/anthropic/messages_transformation.py

@@ -8,6 +8,9 @@
     AnthropicMessagesConfig,
 )
 from litellm.llms.azure.common_utils import BaseAzureLLM
+from litellm.llms.azure_ai.anthropic.output_params_utils import (
+    sanitize_azure_anthropic_output_params,
+)
 from litellm.types.router import GenericLiteLLMParams
 
 
@@ -162,4 +165,5 @@ def transform_anthropic_messages_request(
         )
         self._normalize_system_role_messages(anthropic_messages_request, model=model)
         self._remove_scope_from_cache_control(anthropic_messages_request)
+        sanitize_azure_anthropic_output_params(anthropic_messages_request, model)
         return anthropic_messages_request

litellm/llms/azure_ai/anthropic/output_params_utils.py

@@ -0,0 +1,63 @@
+"""
+Shared sanitization for ``output_config`` on Azure AI Foundry Claude requests.
+Lives in its own module so both the chat-completion transformation
+(``transformation.py``) and the Messages pass-through transformation
+(``messages_transformation.py``) can import it without forming a cycle.
+"""
+
+from typing import Final
+
+
+def _model_accepts_output_config_effort(model: str) -> bool:
+    """Whether ``model`` accepts ``output_config.effort`` on Azure AI Foundry.
+
+    Models that advertise ``supports_output_config`` (or any reasoning effort
+    level) in the model map accept it; others (e.g. Haiku 4.5) return a 400
+    "This model does not support the effort parameter." Imported lazily so
+    this stays a leaf module.
+    """
+    from litellm.llms.anthropic.chat.transformation import AnthropicConfig
+
+    return AnthropicConfig._model_supports_effort_param(model, "azure_ai")
+
+
+def sanitize_azure_anthropic_output_params(
+    data: dict,  # mutable-ok: request dict mutated in place, matches transform_request's contract
+    model: str,
+) -> None:
+    """Strip Azure-unsupported keys from ``output_config`` in-place.
+
+    Behavior:
+      * ``output_config.effort`` is dropped for models that do not accept it
+        (e.g. Haiku 4.5) and forwarded for those that do (Opus/Sonnet 4.6+).
+      * ``output_config`` left empty after filtering is removed so the request
+        body carries no empty dict.
+      * Non-dict values for ``output_config`` are dropped to avoid sending
+        malformed payloads downstream.
+    """
+    output_config: Final = data.get("output_config")
+    if output_config is None:
+        return
+    if not isinstance(output_config, dict):
+        data.pop("output_config", None)
+        return
+
+    drop_effort: Final = "effort" in output_config and not _model_accepts_output_config_effort(model)
+    if drop_effort:
+        from litellm._logging import verbose_logger
+
+        verbose_logger.debug(
+            "Dropping unsupported output_config.effort for azure_ai model=%s "
+            "(no supports_output_config in the model map)",
+            model,
+        )
+
+    sanitized: Final = (
+        {k: v for k, v in output_config.items() if k != "effort"}  # mutable-ok: request dict mutated in place
+        if drop_effort
+        else output_config
+    )
+    if sanitized:
+        data["output_config"] = sanitized  # rebind-ok: out-param store like siblings
+    else:
+        data.pop("output_config", None)

litellm/llms/azure_ai/anthropic/transformation.py

@@ -6,6 +6,9 @@
 
 from litellm.llms.anthropic.chat.transformation import AnthropicConfig
 from litellm.llms.azure.common_utils import BaseAzureLLM
+from litellm.llms.azure_ai.anthropic.output_params_utils import (
+    sanitize_azure_anthropic_output_params,
+)
 from litellm.types.llms.openai import AllMessageValues
 from litellm.types.router import GenericLiteLLMParams
 
@@ -132,4 +135,6 @@ def transform_request(
         data.pop("max_retries", None)
         data.pop("stream_options", None)
 
+        sanitize_azure_anthropic_output_params(data, model)
+
         return data

tests/test_litellm/llms/azure_ai/claude/test_azure_anthropic_messages_transformation.py

@@ -271,6 +271,61 @@ def test_transform_anthropic_messages_request_removes_scope_from_cache_control(
             result["messages"][0]["content"][0]["cache_control"]["type"] == "ephemeral"
         )
 
+    def test_transform_anthropic_messages_request_strips_output_config_effort_for_unsupported_model(
+        self,
+    ):
+        """Regression: output_config.effort forwarded via /anthropic/v1/messages to a model without supports_output_config causes 400 on Azure AI Foundry.
+
+        additional_drop_params does not suppress this because it operates on the
+        OpenAI->provider translation layer, not on the Anthropic pass-through path.
+        See: https://github.com/BerriAI/litellm/issues/27168
+        """
+        config = AzureAnthropicMessagesConfig()
+        model = "claude-haiku-4-5"
+        messages = [{"role": "user", "content": "Hello"}]
+        anthropic_messages_optional_request_params = {
+            "max_tokens": 1024,
+            "output_config": {"effort": "medium"},  # Should be stripped
+        }
+        litellm_params = GenericLiteLLMParams()
+        headers = {}
+
+        result = config.transform_anthropic_messages_request(
+            model=model,
+            messages=messages,
+            anthropic_messages_optional_request_params=anthropic_messages_optional_request_params,
+            litellm_params=litellm_params,
+            headers=headers,
+        )
+
+        assert "output_config" not in result
+        assert result["max_tokens"] == 1024
+
+    def test_transform_anthropic_messages_request_preserves_output_config_for_supported_model(
+        self,
+    ):
+        """output_config must be forwarded for models that advertise supports_output_config (e.g. Sonnet 4.6+)."""
+        config = AzureAnthropicMessagesConfig()
+        model = "claude-sonnet-4-6"
+        messages = [{"role": "user", "content": "Hello"}]
+        anthropic_messages_optional_request_params = {
+            "max_tokens": 1024,
+            "output_config": {"effort": "medium"},
+        }
+        litellm_params = GenericLiteLLMParams()
+        headers = {}
+
+        result = config.transform_anthropic_messages_request(
+            model=model,
+            messages=messages,
+            anthropic_messages_optional_request_params=anthropic_messages_optional_request_params,
+            litellm_params=litellm_params,
+            headers=headers,
+        )
+
+        assert "output_config" in result
+        assert result["output_config"] == {"effort": "medium"}
+
 
 class TestProviderConfigManagerAzureAnthropicMessages:
     """Test ProviderConfigManager returns correct config for Azure AI Anthropic Messages API"""

tests/test_litellm/llms/azure_ai/claude/test_azure_anthropic_output_params_utils.py

@@ -0,0 +1,56 @@
+import os
+import sys
+
+sys.path.insert(
+    0, os.path.abspath(os.path.join(os.path.dirname(__file__), "../../../../.."))
+)
+
+import pytest
+
+from litellm.llms.azure_ai.anthropic.output_params_utils import (
+    sanitize_azure_anthropic_output_params,
+)
+
+
+class TestSanitizeAzureAnthropicOutputParams:
+    def test_drops_effort_for_model_without_supports_output_config(self):
+        """Regression: Haiku 4.5 has no supports_output_config in the model map; effort must be stripped.
+
+        See: https://github.com/BerriAI/litellm/issues/27168
+        """
+        data = {"output_config": {"effort": "medium"}, "max_tokens": 100}
+        sanitize_azure_anthropic_output_params(data, "claude-haiku-4-5")
+        assert "output_config" not in data
+
+    def test_preserves_effort_for_model_with_supports_output_config(self):
+        """Models with supports_output_config (e.g. Sonnet 4.6) must forward effort unchanged."""
+        data = {"output_config": {"effort": "medium"}, "max_tokens": 100}
+        sanitize_azure_anthropic_output_params(data, "claude-sonnet-4-6")
+        assert data["output_config"] == {"effort": "medium"}
+
+    def test_preserves_other_output_config_keys_when_effort_dropped(self):
+        """Only effort is removed; other output_config keys are forwarded."""
+        data = {"output_config": {"effort": "medium", "other_key": "value"}}
+        sanitize_azure_anthropic_output_params(data, "claude-haiku-4-5")
+        assert data["output_config"] == {"other_key": "value"}
+
+    def test_no_output_config_is_noop(self):
+        data: dict = {"max_tokens": 100}
+        sanitize_azure_anthropic_output_params(data, "claude-haiku-4-5")
+        assert "output_config" not in data
+        assert data["max_tokens"] == 100
+
+    def test_non_dict_output_config_is_dropped(self):
+        data = {"output_config": "bad_value"}
+        sanitize_azure_anthropic_output_params(data, "claude-haiku-4-5")
+        assert "output_config" not in data
+
+    def test_empty_output_config_after_effort_drop_is_removed(self):
+        data = {"output_config": {"effort": "high"}}
+        sanitize_azure_anthropic_output_params(data, "claude-haiku-4-5")
+        assert "output_config" not in data
+
+    def test_output_config_without_effort_is_unchanged(self):
+        data = {"output_config": {"other_key": "value"}}
+        sanitize_azure_anthropic_output_params(data, "claude-haiku-4-5")
+        assert data["output_config"] == {"other_key": "value"}

tests/test_litellm/llms/azure_ai/claude/test_azure_anthropic_transformation.py

@@ -242,6 +242,74 @@ def test_transform_request_removes_unsupported_params(self):
         assert result["max_tokens"] == 100
         assert "messages" in result
 
+    def test_transform_request_strips_output_config_effort_for_unsupported_model(self):
+        """Regression test: Azure AI Foundry returns 400 when output_config.effort is forwarded to a model without supports_output_config in the model map (e.g. Haiku 4.5).
+
+        See: https://github.com/BerriAI/litellm/issues/27168
+        """
+        config = AzureAnthropicConfig()
+
+        messages = [{"role": "user", "content": "Hello"}]
+        optional_params = {"max_tokens": 100}
+        litellm_params = {"api_key": "test-key"}
+        headers = {"api-key": "test-key", "anthropic-version": "2023-06-01"}
+
+        with patch.object(
+            config.__class__.__bases__[0],  # AnthropicConfig
+            "transform_request",
+            return_value={
+                "model": "claude-haiku-4-5",
+                "messages": [
+                    {"role": "user", "content": [{"type": "text", "text": "Hello"}]}
+                ],
+                "max_tokens": 100,
+                "output_config": {"effort": "medium"},  # Should be stripped
+            },
+        ):
+            result = config.transform_request(
+                model="claude-haiku-4-5",
+                messages=messages,
+                optional_params=optional_params,
+                litellm_params=litellm_params,
+                headers=headers,
+            )
+
+        assert "output_config" not in result
+        assert result["model"] == "claude-haiku-4-5"
+        assert result["max_tokens"] == 100
+
+    def test_transform_request_preserves_output_config_for_supported_model(self):
+        """output_config must be forwarded for models that advertise supports_output_config (e.g. Sonnet 4.6+)."""
+        config = AzureAnthropicConfig()
+
+        messages = [{"role": "user", "content": "Hello"}]
+        optional_params = {"max_tokens": 100}
+        litellm_params = {"api_key": "test-key"}
+        headers = {"api-key": "test-key", "anthropic-version": "2023-06-01"}
+
+        with patch.object(
+            config.__class__.__bases__[0],  # AnthropicConfig
+            "transform_request",
+            return_value={
+                "model": "claude-sonnet-4-6",
+                "messages": [
+                    {"role": "user", "content": [{"type": "text", "text": "Hello"}]}
+                ],
+                "max_tokens": 100,
+                "output_config": {"effort": "medium"},
+            },
+        ):
+            result = config.transform_request(
+                model="claude-sonnet-4-6",
+                messages=messages,
+                optional_params=optional_params,
+                litellm_params=litellm_params,
+                headers=headers,
+            )
+
+        assert "output_config" in result
+        assert result["output_config"] == {"effort": "medium"}
+
     def test_context_management_compact_beta_header(self):
         """Test that context_management with compact adds the correct beta header for Azure AI"""
         config = AzureAnthropicConfig()