diff --git a/litellm/main.py b/litellm/main.py
index a0454b33736..df3fdef0ea1 100644
--- a/litellm/main.py
+++ b/litellm/main.py
@@ -1284,12 +1284,15 @@ def _get_router_deployment_id(kwargs: Mapping[str, object]) -> str | None:
 
 
 def router_deployment_mode(kwargs: Mapping[str, object]) -> str | None:
-    """The ``model_info.mode`` of the deployment the router picked for this request, if any."""
-    for deployment_model_info in _router_deployment_model_infos(kwargs):
-        mode = deployment_model_info.get("mode")
-        if isinstance(mode, str):
-            return mode
-    return None
+    """The ``mode`` the router registered for the deployment it picked for this request, if any.
+
+    Only the deployment id is taken from request metadata; the mode itself is read from the
+    router's own registration, so a caller can't hand-write one into their metadata.
+    """
+    deployment_id: Final = _get_router_deployment_id(kwargs)
+    registered: Final = litellm.model_cost.get(deployment_id) if deployment_id is not None else None
+    mode: Final = registered.get("mode") if isinstance(registered, dict) else None
+    return mode if isinstance(mode, str) else None
 
 
 def _register_custom_pricing_for_request(
diff --git a/tests/unit/llms/anthropic/pass_through/adapters/test_handler_reasoning_effort_normalization.py b/tests/unit/llms/anthropic/pass_through/adapters/test_handler_reasoning_effort_normalization.py
index 1d8ebb997c8..32b048e43a4 100644
--- a/tests/unit/llms/anthropic/pass_through/adapters/test_handler_reasoning_effort_normalization.py
+++ b/tests/unit/llms/anthropic/pass_through/adapters/test_handler_reasoning_effort_normalization.py
@@ -10,6 +10,7 @@ from typing import Final
 
 import pytest
 
+import litellm
 from litellm.llms.anthropic.pass_through.adapters.handler import (
     LiteLLMMessagesToCompletionTransformationHandler,
 )
@@ -122,11 +123,12 @@ class TestTheSummaryWrappingOnlyRidesTheResponsesBridge:
         [("responses", {"effort": "high", "summary": "auto"}), ("chat", "high")],
     )
     def test_the_routed_deployments_own_mode_decides_the_shape(
-        self, local_model_cost_map: None, deployment_mode: str, expected: object
+        self, local_model_cost_map: None, monkeypatch: pytest.MonkeyPatch, deployment_mode: str, expected: object
     ) -> None:
-        """https://github.com/BerriAI/litellm/issues/38543 - a deployment's mode rides with the
-        request instead of the model's shared cost-map entry, and this probe has to agree with
+        """https://github.com/BerriAI/litellm/issues/38543 - the routed deployment's own mode, not the
+        model's shared cost-map entry, decides the API, and this probe has to agree with
         ``litellm.completion`` about which API that request lands on"""
+        monkeypatch.setitem(litellm.model_cost, "routed-deployment", {"mode": deployment_mode})
         completion_kwargs, _ = LiteLLMMessagesToCompletionTransformationHandler._prepare_completion_kwargs(
             max_tokens=1024,
             messages=MESSAGES,
@@ -144,7 +146,7 @@ class TestTheSummaryWrappingOnlyRidesTheResponsesBridge:
             output_format=None,
             extra_kwargs={
                 "custom_llm_provider": "databricks",
-                "litellm_metadata": {"model_info": {"id": "routed-deployment", "mode": deployment_mode}},
+                "litellm_metadata": {"model_info": {"id": "routed-deployment"}},
             },
         )
 
diff --git a/tests/unit/responses/litellm_completion_transformation/test_litellm_completion_responses.py b/tests/unit/responses/litellm_completion_transformation/test_litellm_completion_responses.py
index 0ea38077e99..3030620f26a 100644
--- a/tests/unit/responses/litellm_completion_transformation/test_litellm_completion_responses.py
+++ b/tests/unit/responses/litellm_completion_transformation/test_litellm_completion_responses.py
@@ -2770,19 +2770,20 @@ class TestToolTransformation:
         assert result["reasoning_summary"] == "auto"
 
     @pytest.mark.parametrize("deployment_mode, expected_summary", [("responses", "auto"), ("chat", None)])
-    def test_summary_follows_the_routed_deployments_own_mode(self, deployment_mode, expected_summary):
+    def test_summary_follows_the_routed_deployments_own_mode(self, monkeypatch, deployment_mode, expected_summary):
         """
         https://github.com/BerriAI/litellm/issues/38543 - a deployment's ``mode`` no longer lands
         on the model's shared cost-map entry, so the probe has to read it off the routed
         deployment the same way ``litellm.completion`` does, or it drops the summary of a
         request that completion then bridges onto the Responses API
         """
+        monkeypatch.setitem(litellm.model_cost, "routed-deployment", {"mode": deployment_mode})
         result = LiteLLMCompletionResponsesConfig.transform_responses_api_request_to_chat_completion_request(
             model="gpt-4.1",
             input="hi",
             responses_api_request={"reasoning": {"effort": "medium", "summary": "auto"}},
             custom_llm_provider="openai",
-            litellm_metadata={"model_info": {"id": "routed-deployment", "mode": deployment_mode}},
+            litellm_metadata={"model_info": {"id": "routed-deployment"}},
         )
 
         assert result["reasoning_effort"] == "medium"
diff --git a/tests/unit/test_router_model_cost_isolation.py b/tests/unit/test_router_model_cost_isolation.py
index 67152cf7c92..a4c4d8ddf18 100644
--- a/tests/unit/test_router_model_cost_isolation.py
+++ b/tests/unit/test_router_model_cost_isolation.py
@@ -854,6 +854,36 @@ def test_a_priced_responses_deployment_does_not_move_its_siblings_after_serving_
     _invalidate_model_cost_lowercase_map()
 
 
+@pytest.mark.parametrize(
+    "caller_model_info",
+    ({"mode": "responses"}, {"id": "not-a-deployment", "mode": "responses"}),
+)
+def test_a_caller_cannot_move_a_chat_deployment_onto_the_responses_api_through_its_metadata(
+    respx_mock, monkeypatch: pytest.MonkeyPatch, caller_model_info: Mapping[str, str]
+) -> None:
+    """
+    The Responses bridge reads the mode the router registered for the routed deployment, never a
+    mode written into the request's own metadata
+    """
+    _use_catalog_with_a_chat_model(monkeypatch)
+    chat_route: Final = respx_mock.post(f"{_MODE_TEST_API_BASE}/chat/completions").mock(
+        return_value=httpx.Response(200, json=_CHAT_COMPLETION_REPLY)
+    )
+    responses_route: Final = respx_mock.post(f"{_MODE_TEST_API_BASE}/responses").mock(
+        return_value=httpx.Response(200, json=_RESPONSES_REPLY)
+    )
+    router: Final = Router(model_list=[_mode_test_deployment("my-chat-model", None)])
+
+    router.completion(
+        model="my-chat-model",
+        messages=[{"role": "user", "content": "hi"}],
+        litellm_metadata={"model_info": dict(caller_model_info)},
+    )
+
+    assert (chat_route.call_count, responses_route.call_count) == (1, 0)
+    _invalidate_model_cost_lowercase_map()
+
+
 def test_deployment_mode_still_fills_a_shared_key_with_no_mode(monkeypatch: pytest.MonkeyPatch) -> None:
     """
     For a provider model the catalog doesn't know, proxy model_info is how its mode gets