diff --git a/litellm/proxy/proxy_server.py b/litellm/proxy/proxy_server.py
index 10c508fccaa..85e9bdd05e6 100644
--- a/litellm/proxy/proxy_server.py
+++ b/litellm/proxy/proxy_server.py
@@ -14621,13 +14621,6 @@ def _apply_webhook_role_gate(webhook_map, is_full_admin: bool):
 
 
 def _apply_router_settings_role_gate(router_settings: dict, is_full_admin: bool) -> dict:
-    """
-    ``default_litellm_params`` is an open-ended kwargs dict (Router merges it into
-    every completion call), so an admin can put a shared ``api_key`` or an
-    ``extra_headers`` Authorization token there. Unlike the other fields on
-    this page, nothing else masks it before it reaches non-full-admin callers
-    of /get/config/callbacks.
-    """
     default_litellm_params = router_settings.get("default_litellm_params")
     if is_full_admin or not isinstance(default_litellm_params, dict):
         return router_settings
diff --git a/litellm/router.py b/litellm/router.py
index 58427a1797f..18f7f79f965 100644
--- a/litellm/router.py
+++ b/litellm/router.py
@@ -9706,23 +9706,11 @@ class Router:
         _settings_to_return["routing_groups"] = [group.model_dump() for group in self._routing_groups.values()]
         return _settings_to_return
 
-    def _merge_default_litellm_params_setting(self, value: dict | None) -> None:
+    def _replace_default_litellm_params_setting(self, value: dict | None) -> None:
         if value is not None:
-            self.default_litellm_params = {**self.default_litellm_params, **value}
+            self.default_litellm_params = value
 
     def _remove_optional_pre_call_checks(self, removed_checks: OptionalPreCallChecks) -> list[str]:
-        """
-        Reverse of `add_optional_pre_call_checks` for the subset of checks that can be
-        safely turned off at runtime: clears the corresponding flag(s) on shared
-        affinity callbacks, or unregisters the dedicated callback instance entirely.
-
-        Returns the subset of `removed_checks` that were kept active anyway because
-        they're still required by config (currently only `router_budget_limiting`,
-        which `Router.__init__` auto-enables whenever budgets are configured on the
-        deployments/provider, independent of `optional_pre_call_checks`) - callers
-        should fold these back into the tracked `optional_pre_call_checks` list so it
-        doesn't claim a check is off when it's actually still enforced.
-        """
         from litellm.router_utils.pre_call_checks.encrypted_content_affinity_check import (
             EncryptedContentAffinityCheck,
         )
@@ -9762,9 +9750,6 @@ class Router:
             if check == "router_budget_limiting" and RouterBudgetLimiting.should_init_router_budget_limiter(
                 model_list=self.model_list, provider_budget_config=self.provider_budget_config
             ):
-                # Budgets are still configured on the deployments/provider - Router.__init__
-                # would auto-enable this regardless of optional_pre_call_checks, so a save
-                # that omits it must not silently disable budget enforcement.
                 retained_checks.append(check)
                 continue
             litellm.logging_callback_manager.remove_callbacks_by_type(optional_callbacks, callback_type)
@@ -9783,12 +9768,8 @@ class Router:
         retained_checks = self._remove_optional_pre_call_checks(removed_checks) if removed_checks else []
         self.optional_pre_call_checks = list(dict.fromkeys([*value, *retained_checks]))
 
-    # Settings whose update logic doesn't fit `setattr(self, var, value)` (e.g.
-    # merge-not-replace, or side effects beyond storing the value). Dispatched via
-    # a single `var in ...` branch in update_settings so adding an entry here
-    # doesn't grow that function's branch count per field.
     _CUSTOM_UPDATE_SETTINGS_HANDLERS: dict = {
-        "default_litellm_params": _merge_default_litellm_params_setting,
+        "default_litellm_params": _replace_default_litellm_params_setting,
         "optional_pre_call_checks": _apply_optional_pre_call_checks_setting,
     }
 
diff --git a/litellm/types/management_endpoints/router_settings_endpoints.py b/litellm/types/management_endpoints/router_settings_endpoints.py
index 72d1cd2f30d..fb3d6e10fcf 100644
--- a/litellm/types/management_endpoints/router_settings_endpoints.py
+++ b/litellm/types/management_endpoints/router_settings_endpoints.py
@@ -243,9 +243,6 @@ ROUTER_SETTINGS_FIELDS: List[RouterSettingsField] = [
             "that cached the prompt."
         ),
         field_default=[],
-        # "forward_client_headers_by_model_group" is excluded: it's a literal in the
-        # OptionalPreCallChecks type union, but add_optional_pre_call_checks has no
-        # handler for it, so selecting it from the UI would silently do nothing.
         options=[
             "prompt_caching",
             "router_budget_limiting",
diff --git a/tests/code_coverage_tests/router_code_coverage.py b/tests/code_coverage_tests/router_code_coverage.py
index 9ee210c0bb2..fcd1ca2f392 100644
--- a/tests/code_coverage_tests/router_code_coverage.py
+++ b/tests/code_coverage_tests/router_code_coverage.py
@@ -80,7 +80,7 @@ ignored_function_names = [
     "_override_vector_store_methods_for_router",  # No-op placeholder, called during Router init
     "_merge_tools_from_deployment",  # Tested indirectly via _update_kwargs_with_deployment (test files lack "router" in name)
     "_invalidate_access_groups_cache",  # Tested indirectly via set_model_list, upsert_model etc. (test files lack "router" in name)
-    "_merge_default_litellm_params_setting",  # Tested indirectly via update_settings(default_litellm_params=...)
+    "_replace_default_litellm_params_setting",  # Tested indirectly via update_settings(default_litellm_params=...)
     "_apply_optional_pre_call_checks_setting",  # Tested indirectly via update_settings(optional_pre_call_checks=...)
     "_remove_optional_pre_call_checks",  # Tested indirectly via update_settings(optional_pre_call_checks=...)
 ]
diff --git a/tests/test_litellm/test_router.py b/tests/test_litellm/test_router.py
index fcd2725c185..57a4bca46fe 100644
--- a/tests/test_litellm/test_router.py
+++ b/tests/test_litellm/test_router.py
@@ -5065,19 +5065,15 @@ def _make_router_for_settings_tests(**kwargs):
     )
 
 
-def test_update_settings_merges_default_litellm_params_without_dropping_existing_keys():
-    """
-    Regression test: `update_settings(default_litellm_params=...)` must merge into
-    the existing dict, not replace it wholesale. A naive `setattr` replace would
-    silently drop keys the Router set at init (e.g. `timeout`, `max_retries`,
-    `metadata`) whenever an admin edits `default_litellm_params` from the UI to
-    add something like `cache_control_injection_points`.
-    """
+def test_update_settings_replaces_default_litellm_params_wholesale():
     router = _make_router_for_settings_tests(timeout=42)
     assert router.default_litellm_params.get("timeout") == 42
 
     router.update_settings(
-        default_litellm_params={"cache_control_injection_points": [{"location": "message", "role": "system"}]}
+        default_litellm_params={
+            "timeout": 42,
+            "cache_control_injection_points": [{"location": "message", "role": "system"}],
+        }
     )
 
     assert router.default_litellm_params["timeout"] == 42
@@ -5086,6 +5082,25 @@ def test_update_settings_merges_default_litellm_params_without_dropping_existing
     ]
 
 
+def test_update_settings_can_clear_default_litellm_params_keys():
+    """
+    Regression test: a prior merge-not-replace implementation could never remove a
+    key - `{**old, **new}` keeps any key present in `old` but absent from `new`.
+    Saving `default_litellm_params` after removing `cache_control_injection_points`
+    from the Admin UI's textarea must actually clear it, not silently keep the old
+    callback-affecting value active.
+    """
+    router = _make_router_for_settings_tests()
+    router.update_settings(
+        default_litellm_params={"cache_control_injection_points": [{"location": "message", "role": "system"}]}
+    )
+    assert "cache_control_injection_points" in router.default_litellm_params
+
+    router.update_settings(default_litellm_params={})
+
+    assert "cache_control_injection_points" not in router.default_litellm_params
+
+
 def test_update_settings_optional_pre_call_checks_is_idempotent():
     """
     Regression test: `add_optional_pre_call_checks` has no built-in guard against
diff --git a/ui/litellm-dashboard/src/components/router_settings/index.test.tsx b/ui/litellm-dashboard/src/components/router_settings/index.test.tsx
index 536548a95b9..e155fbddf84 100644
--- a/ui/litellm-dashboard/src/components/router_settings/index.test.tsx
+++ b/ui/litellm-dashboard/src/components/router_settings/index.test.tsx
@@ -207,11 +207,6 @@ describe("RouterSettings", () => {
   });
 
   it("should round-trip default_litellm_params and optional_pre_call_checks unmodified on save", async () => {
-    // Regression test: default_litellm_params holds a dict (e.g. cache_control_injection_points),
-    // not a plain string - without listing it in the save handler's jsonKeys set, it'd be persisted
-    // as raw stringified text instead of parsed JSON. optional_pre_call_checks is a list owned by
-    // its own multi-select (no DOM input); without the fallback-to-state path in parseInputValue,
-    // an untouched value would be silently dropped instead of round-tripping through Save.
     vi.mocked(getCallbacksCall).mockResolvedValue({
       router_settings: {
         ...mockCallbacksResponse.router_settings,