diff --git a/litellm/litellm_core_utils/model_param_helper.py b/litellm/litellm_core_utils/model_param_helper.py
index b4fa5cb60aa..855e52098c4 100644
--- a/litellm/litellm_core_utils/model_param_helper.py
+++ b/litellm/litellm_core_utils/model_param_helper.py
@@ -1,3 +1,4 @@
+from functools import lru_cache
 from typing import Set
 
 from openai.types.chat.completion_create_params import (
@@ -54,9 +55,17 @@ class ModelParamHelper:
         return combined_kwargs
 
     @staticmethod
+    @lru_cache(maxsize=1)
     def _get_all_llm_api_params() -> Set[str]:
         """
-        Gets the supported kwargs for each call type and combines them
+        Gets the supported kwargs for each call type and combines them.
+
+        The result is derived from static type annotations and fixed sets, so it
+        is constant for the process lifetime. It is computed once and cached
+        because it is rebuilt on every request through both the cache-key path
+        (``Cache.get_cache_key``) and the spend-logging path
+        (``_get_relevant_args_to_use_for_logging``). Callers treat the result as
+        read-only.
         """
         chat_completion_kwargs = (
             ModelParamHelper._get_litellm_supported_chat_completion_kwargs()
diff --git a/tests/test_litellm/litellm_core_utils/test_model_param_helper.py b/tests/test_litellm/litellm_core_utils/test_model_param_helper.py
new file mode 100644
index 00000000000..df01bd636b8
--- /dev/null
+++ b/tests/test_litellm/litellm_core_utils/test_model_param_helper.py
@@ -0,0 +1,26 @@
+import os
+import sys
+
+sys.path.insert(
+    0, os.path.abspath("../../..")
+)  # Adds the parent directory to the system path
+
+from litellm.litellm_core_utils.model_param_helper import ModelParamHelper
+
+
+def test_get_all_llm_api_params_is_correct():
+    """The cached result must equal a fresh, uncached computation."""
+    cached = ModelParamHelper._get_all_llm_api_params()
+    uncached = ModelParamHelper._get_all_llm_api_params.__wrapped__()
+    assert cached == uncached
+    assert {"model", "temperature", "stream"} <= cached
+    assert "metadata" not in cached  # excluded via _get_exclude_kwargs
+
+
+def test_get_all_llm_api_params_is_memoized():
+    """Regression: the param set is static and is rebuilt on every request via
+    the cache-key and spend-logging paths, so it must be memoized. Without the
+    cache each call returns a freshly built set (a different object)."""
+    first = ModelParamHelper._get_all_llm_api_params()
+    second = ModelParamHelper._get_all_llm_api_params()
+    assert first is second