fix(anthropic): gate sampling params on /v1/messages by supports_sampling_params

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

@@ -1,12 +1,16 @@
 from functools import lru_cache
 from typing import Any, Dict, FrozenSet, List, cast, get_type_hints
 
+from litellm.llms.anthropic.common_utils import AnthropicModelInfo
 from litellm.types.llms.anthropic import AnthropicMessagesRequestOptionalParams
 from litellm.types.llms.anthropic_messages.anthropic_response import (
     AnthropicMessagesResponse,
 )
 
 
+_SAMPLING_PARAMS: frozenset[str] = frozenset(("temperature", "top_p", "top_k"))
+
+
 @lru_cache(maxsize=1)
 def _anthropic_messages_optional_param_keys() -> FrozenSet[str]:
     """
@@ -51,6 +55,18 @@ class AnthropicMessagesRequestUtils:
                 drop_params=drop_params,
                 custom_llm_provider=custom_llm_provider,
             )
+            for param in _SAMPLING_PARAMS:
+                if param not in filtered_params:
+                    continue
+                value = filtered_params.pop(param)
+                AnthropicModelInfo._apply_sampling_param(
+                    optional_params=filtered_params,
+                    model=model,
+                    param=param,
+                    value=value,
+                    drop_params=drop_params,
+                    output_key=param,
+                )
         return cast(AnthropicMessagesRequestOptionalParams, filtered_params)
 
 

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

@@ -6,6 +6,8 @@ Regression tests for the /v1/messages request-parse fast paths:
   while resolving the (static) type hints only once per process.
 """
 
+import pytest
+
 import litellm
 from litellm.llms.anthropic.experimental_pass_through.messages.utils import (
     AnthropicMessagesRequestUtils,
@@ -88,3 +90,48 @@ def test_drop_params_keeps_speed_for_supporting_model():
         litellm.drop_params = original
 
     assert result == {"speed": "fast"}
+
+
+@pytest.mark.parametrize("param,value", [("temperature", 0.3), ("top_p", 0.9), ("top_k", 40)])
+def test_sampling_params_dropped_for_models_without_sampling_support(param, value):
+    original = litellm.drop_params
+    litellm.drop_params = True
+    try:
+        result = AnthropicMessagesRequestUtils.get_requested_anthropic_messages_optional_param(
+            params={param: value, "max_tokens": 16},
+            model="claude-opus-4-8",
+        )
+    finally:
+        litellm.drop_params = original
+
+    assert result == {"max_tokens": 16}
+
+
+@pytest.mark.parametrize("param,value", [("temperature", 0.3), ("top_p", 0.9), ("top_k", 40)])
+def test_sampling_params_raise_without_drop_params(param, value):
+    original = litellm.drop_params
+    litellm.drop_params = False
+    try:
+        with pytest.raises(litellm.utils.UnsupportedParamsError):
+            AnthropicMessagesRequestUtils.get_requested_anthropic_messages_optional_param(
+                params={param: value},
+                model="claude-opus-4-8",
+            )
+    finally:
+        litellm.drop_params = original
+
+
+def test_sampling_params_kept_for_supporting_model():
+    result = AnthropicMessagesRequestUtils.get_requested_anthropic_messages_optional_param(
+        params={"temperature": 0.3, "top_p": 0.9, "top_k": 40},
+        model="claude-sonnet-4-6",
+    )
+    assert result == {"temperature": 0.3, "top_p": 0.9, "top_k": 40}
+
+
+def test_temperature_one_allowed_for_model_without_sampling_support():
+    result = AnthropicMessagesRequestUtils.get_requested_anthropic_messages_optional_param(
+        params={"temperature": 1},
+        model="claude-opus-4-8",
+    )
+    assert result == {"temperature": 1}