diff --git a/litellm/router.py b/litellm/router.py
index 6a43a6a05c0..474c284c389 100644
--- a/litellm/router.py
+++ b/litellm/router.py
@@ -2857,6 +2857,17 @@ class Router:
                             MID_STREAM_CONTINUATION_MARKER,
                         )
 
+                        emitted_tokens: Final = int(
+                            getattr(complete_response_object_usage, "completion_tokens", 0) or 0
+                        )
+                        reduced_ceilings: Final = self._continuation_output_ceilings(initial_kwargs, emitted_tokens)
+                        if reduced_ceilings is None:
+                            # The caller's output allowance is already spent; surface the
+                            # error rather than grant a fresh allowance on this fallback hop.
+                            if e.original_exception is not None:
+                                raise e.original_exception from e
+                            raise
+                        initial_kwargs.update(reduced_ceilings)
                         initial_kwargs["messages"] = self._build_completion_continuation_input(
                             messages, e.generated_content
                         )
@@ -3076,6 +3087,28 @@ class Router:
             return False
         return True
 
+    @staticmethod
+    def _continuation_output_ceilings(
+        request_kwargs: Mapping[str, object],
+        emitted_tokens: int,
+    ) -> Mapping[str, int] | None:
+        """The max_tokens / max_completion_tokens a continuation should carry, each
+        reduced by the tokens already emitted so the whole answer stays within the
+        caller's original allowance instead of getting a fresh one on every fallback
+        hop. Reductions compound across hops because the trimmed ceiling is what the
+        next hop sees. Returns None when the allowance is already exhausted, so the
+        stream must not be continued."""
+        ceilings: Final = MappingProxyType(
+            {
+                key: value - emitted_tokens
+                for key in ("max_tokens", "max_completion_tokens")
+                if isinstance(value := request_kwargs.get(key), int)
+            }
+        )
+        if ceilings and min(ceilings.values()) <= 0:
+            return None
+        return ceilings
+
     @staticmethod
     def _build_completion_continuation_input(
         messages: Sequence[Mapping[str, str]],
diff --git a/litellm/router_utils/pre_call_checks/continuation_prefill_check.py b/litellm/router_utils/pre_call_checks/continuation_prefill_check.py
index ccd454bc207..c640374858a 100644
--- a/litellm/router_utils/pre_call_checks/continuation_prefill_check.py
+++ b/litellm/router_utils/pre_call_checks/continuation_prefill_check.py
@@ -46,6 +46,7 @@ def _deployment_supports_prefill(deployment: object) -> bool:
         if isinstance(declared, bool):
             return declared
     except ValidationError:
+        # No usable model_info override; fall through to the cost-map lookup.
         pass
     try:
         litellm_params: Final = _STR_KEYED_DICT_ADAPTER.validate_python(deployment_map.get("litellm_params"))
diff --git a/tests/test_litellm/test_router.py b/tests/test_litellm/test_router.py
index c37a23df7d1..987b96b3228 100644
--- a/tests/test_litellm/test_router.py
+++ b/tests/test_litellm/test_router.py
@@ -2363,6 +2363,26 @@ def test_build_completion_continuation_input_folds_into_existing_prefill():
     assert twice[-1] == {"role": "assistant", "content": "part one part two", "prefix": True}
 
 
+def test_continuation_output_ceilings_reduces_by_emitted_tokens():
+    """A continuation must complete within the caller's original allowance, so each
+    output ceiling is reduced by the tokens already emitted."""
+    assert litellm.Router._continuation_output_ceilings({"max_tokens": 100}, 30) == {"max_tokens": 70}
+    assert litellm.Router._continuation_output_ceilings({"max_tokens": 100, "max_completion_tokens": 40}, 25) == {
+        "max_tokens": 75,
+        "max_completion_tokens": 15,
+    }
+    # no ceiling configured -> nothing to reduce, continuation proceeds as before
+    assert litellm.Router._continuation_output_ceilings({}, 50) == {}
+
+
+def test_continuation_output_ceilings_none_when_allowance_exhausted():
+    """When the emitted tokens already meet or exceed a ceiling, there is no budget
+    left to continue, so the helper signals a decline rather than a fresh allowance."""
+    assert litellm.Router._continuation_output_ceilings({"max_tokens": 20}, 20) is None
+    assert litellm.Router._continuation_output_ceilings({"max_tokens": 20}, 25) is None
+    assert litellm.Router._continuation_output_ceilings({"max_tokens": 100, "max_completion_tokens": 10}, 10) is None
+
+
 def test_mid_stream_continuation_eligible_allows_text_response_format():
     """response_format={"type": "text"} is the unconstrained default and must stay
     eligible, unlike json_object / json_schema."""