diff --git a/litellm/router_utils/reasoning_effort_capability.py b/litellm/router_utils/reasoning_effort_capability.py
index 399f97a9864..9185d901a28 100644
--- a/litellm/router_utils/reasoning_effort_capability.py
+++ b/litellm/router_utils/reasoning_effort_capability.py
@@ -149,17 +149,19 @@ def resolve_supported_reasoning_efforts(
     on. deployment_is_mapped is that provenance, and an operator who wants either answer for an
     off-map deployment gets it by setting supports_reasoning explicitly.
 
-    If no explicit supports_reasoning flag is set but at least one per-level flag (e.g.
-    supports_minimal_reasoning_effort) is present, treat supports_reasoning as implicitly True,
-    since the per-level flags are evidence the model supports reasoning.
+    If supports_reasoning is unset but at least one per-level flag (e.g.
+    supports_minimal_reasoning_effort) is present, treat it as implicitly True, since the
+    per-level flags are evidence the model supports reasoning. An explicit False always wins:
+    it is the operator's escape hatch and must not be overridden by inherited per-level flags.
     """
     supports_reasoning: Final = model_info.get("supports_reasoning")
+    if supports_reasoning is False:
+        return ()
+
     flags: Final = _declared_effort_flags(model_info)
     has_per_level_flag: Final = any(value is not None for value in flags.values())
-
-    if supports_reasoning is not True:
-        if not has_per_level_flag:
-            return () if supports_reasoning is False or deployment_is_mapped else None
+    if supports_reasoning is not True and not has_per_level_flag:
+        return () if deployment_is_mapped else None
 
     declared: Final = declared_reasoning_efforts(model_info)
     if declared is not None:
diff --git a/tests/test_litellm/router_utils/test_reasoning_effort_capability.py b/tests/test_litellm/router_utils/test_reasoning_effort_capability.py
index fd5926b75f0..a0dbf3b6637 100644
--- a/tests/test_litellm/router_utils/test_reasoning_effort_capability.py
+++ b/tests/test_litellm/router_utils/test_reasoning_effort_capability.py
@@ -86,9 +86,6 @@ class TestResolveSupportedReasoningEfforts:
         assert resolved == ("none", "minimal", "low", "medium", "high")
 
     def test_per_level_flag_without_supports_reasoning_treats_as_implicit_true(self):
-        # A model with only a per-level flag (e.g. supports_minimal_reasoning_effort) but no
-        # explicit supports_reasoning should be treated as reasoning-capable, since the per-level
-        # flag is evidence of reasoning support.
         resolved = resolve_supported_reasoning_efforts(
             {
                 "supports_minimal_reasoning_effort": True,
@@ -97,6 +94,16 @@ class TestResolveSupportedReasoningEfforts:
         )
         assert resolved == ("none", "minimal", "low", "medium", "high")
 
+    def test_explicit_supports_reasoning_false_wins_over_per_level_flags(self):
+        resolved = resolve_supported_reasoning_efforts(
+            {
+                "supports_reasoning": False,
+                "supports_minimal_reasoning_effort": True,
+            },
+            deployment_is_mapped=True,
+        )
+        assert resolved == ()
+
 
 class TestBareModelNameFallback:
     def test_a_prefixed_entry_inherits_the_flags_of_its_unprefixed_twin(self):