diff --git a/litellm/integrations/anthropic_cache_control_hook.py b/litellm/integrations/anthropic_cache_control_hook.py
index be4157d3f05..1db144b5fdc 100644
--- a/litellm/integrations/anthropic_cache_control_hook.py
+++ b/litellm/integrations/anthropic_cache_control_hook.py
@@ -695,6 +695,7 @@ class AnthropicCacheControlHook(CustomPromptManagement):
         tools: list | None = None,
         cache_control: object = None,
         request_kwargs: object = None,
+        on_messages_route: bool = False,
     ) -> bool:
         """Return True if the request already carries any client-supplied cache_control.
 
@@ -704,12 +705,14 @@ class AnthropicCacheControlHook(CustomPromptManagement):
         envelope. Configured injection points are an explicit instruction and are
         applied alongside the client's marks, bounded by the provider cap.
         """
-        return (
-            AnthropicCacheControlHook.count_request_cache_breakpoints(messages, system)
-            + AnthropicCacheControlHook.count_external_cache_breakpoints_on_messages_route(
+        external_breakpoints: Final = (
+            AnthropicCacheControlHook.count_external_cache_breakpoints_on_messages_route(
                 tools, cache_control, request_kwargs
             )
-        ) > 0
+            if on_messages_route
+            else AnthropicCacheControlHook.count_external_cache_breakpoints(tools, cache_control, request_kwargs)
+        )
+        return AnthropicCacheControlHook.count_request_cache_breakpoints(messages, system) + external_breakpoints > 0
 
     @staticmethod
     def get_default_injection_points(
@@ -721,6 +724,7 @@ class AnthropicCacheControlHook(CustomPromptManagement):
         enable_prompt_caching: bool | None = None,
         cache_control: object = None,
         request_kwargs: object = None,
+        on_messages_route: bool = False,
     ) -> list[CacheControlInjectionPoint]:
         """Default breakpoints when ``litellm.enable_anthropic_prompt_caching`` is on.
 
@@ -741,7 +745,9 @@ class AnthropicCacheControlHook(CustomPromptManagement):
         if not supports_anthropic_cache_control(model, custom_llm_provider):
             return []
 
-        if AnthropicCacheControlHook._request_has_cache_control(messages, system, tools, cache_control, request_kwargs):
+        if AnthropicCacheControlHook._request_has_cache_control(
+            messages, system, tools, cache_control, request_kwargs, on_messages_route
+        ):
             return []
 
         if is_claude_code_one_shot_subagent_request(
@@ -970,6 +976,7 @@ class AnthropicCacheControlHook(CustomPromptManagement):
                 enable_prompt_caching=enable_prompt_caching,
                 cache_control=cache_control,
                 request_kwargs=kwargs,
+                on_messages_route=True,
             )
             if model is not None
             else ()
diff --git a/tests/unit/integrations/test_anthropic_cache_control_hook.py b/tests/unit/integrations/test_anthropic_cache_control_hook.py
index 147e7423a9d..1d70a21af7b 100644
--- a/tests/unit/integrations/test_anthropic_cache_control_hook.py
+++ b/tests/unit/integrations/test_anthropic_cache_control_hook.py
@@ -2657,6 +2657,18 @@ class TestConfiguredInjectionPointsSurviveClientMarks:
         )
         assert ("litellm_gateway_injected_cache" in request_kwargs["litellm_metadata"]) is injected
 
+    def test_chat_automatic_defaults_apply_when_extra_body_drops_the_only_client_mark(self, monkeypatch):
+        monkeypatch.setattr(litellm, "enable_anthropic_prompt_caching", True)
+        params = {"extra_body": {"tools": [self.UNMARKED_TOOL]}}
+
+        self._seed(params, copy.deepcopy(self.CLEAN_MESSAGES), tools=[self.MARKED_TOOL_TOP_LEVEL])
+        affinity = AnthropicCacheControlHook.messages_with_default_injections(
+            copy.deepcopy(self.CLEAN_MESSAGES), ["claude-sonnet-4-5"], tools=[self.MARKED_TOOL_TOP_LEVEL], request_kwargs=params
+        )
+
+        assert [p["index"] for p in params["cache_control_injection_points"]] == [None, -1]
+        assert AnthropicCacheControlHook.count_request_cache_breakpoints(affinity) == 2
+
     @pytest.mark.parametrize(
         "marked_turns,expected_system",
         [(2, [{"type": "text", "text": "sys", "cache_control": {"type": "ephemeral"}}]), (3, "sys")],