diff --git a/enterprise/litellm_enterprise/proxy/common_utils/check_batch_cost.py b/enterprise/litellm_enterprise/proxy/common_utils/check_batch_cost.py
index 4bb00408fc32..76e92538aaa7 100644
--- a/enterprise/litellm_enterprise/proxy/common_utils/check_batch_cost.py
+++ b/enterprise/litellm_enterprise/proxy/common_utils/check_batch_cost.py
@@ -966,6 +966,16 @@ async def check_batch_cost(self):
                     )
 
             elif response.status in PROVIDER_TERMINAL_BATCH_STATUSES:
+                from litellm.proxy.openai_files_endpoints.common_utils import (
+                    _completed_batch_safe_to_retire,
+                )
+
+                if response.status in ("completed", "complete") and not _completed_batch_safe_to_retire(response):
+                    verbose_proxy_logger.info(
+                        f"CheckBatchCost: batch {batch_id} is completed but its output file id "
+                        f"has not appeared yet; leaving job {job.id} for the next poll cycle"
+                    )
+                    continue
                 await self._finalize_unbilled_terminal_job(job, response)
 
         # Record polling run metrics (always, even if nothing was processed)
diff --git a/litellm/proxy/openai_files_endpoints/common_utils.py b/litellm/proxy/openai_files_endpoints/common_utils.py
index d4cb8fe2374a..142aced4a38f 100644
--- a/litellm/proxy/openai_files_endpoints/common_utils.py
+++ b/litellm/proxy/openai_files_endpoints/common_utils.py
@@ -1334,6 +1334,25 @@ def batch_cost_poller_is_active() -> bool:
         return False
 
 
+def _completed_batch_safe_to_retire(response: "LiteLLMBatch") -> bool:
+    """Whether a "completed" batch may be retired from cost recovery.
+
+    ``batch_processed=True`` is the sole re-pickup gate for CheckBatchCost's
+    cost-recovery poller, so setting it retires the batch permanently. A batch can
+    reach ``status="completed"`` while ``output_file_id`` is still ``None`` (the
+    provider response briefly lags before the output id populates). Retiring in that
+    window loses the spend record forever. Retire only once we can prove there is
+    nothing left to recover: the output file has actually arrived, or the provider
+    reports no successful request lines. When counts are unknown, stay eligible so
+    the next poller pass revisits it. (#37713)
+    """
+    if getattr(response, "output_file_id", None) is not None:
+        return True
+    request_counts = getattr(response, "request_counts", None)
+    completed = getattr(request_counts, "completed", None)
+    return completed == 0
+
+
 async def update_batch_in_database(
     batch_id: str,
     unified_batch_id: str | Literal[False],
@@ -1415,7 +1434,7 @@ async def update_batch_in_database(
         }
 
         poller_owns: Final = batch_cost_poller_is_active() if poller_owns_accounting is None else poller_owns_accounting
-        if db_status == "complete" and not poller_owns:
+        if db_status == "complete" and not poller_owns and _completed_batch_safe_to_retire(response):
             update_data["batch_processed"] = True
 
         try:
diff --git a/tests/proxy_unit_tests/test_check_batch_cost.py b/tests/proxy_unit_tests/test_check_batch_cost.py
index 0065dbebc596..a1864c5e4809 100644
--- a/tests/proxy_unit_tests/test_check_batch_cost.py
+++ b/tests/proxy_unit_tests/test_check_batch_cost.py
@@ -1044,7 +1044,9 @@ async def test_completed_without_output_file_marked_processed_without_billing(
         Pre-fix it matched neither the completed-with-output branch nor the
         failed/expired/cancelled branch, so batch_processed stayed False and the row
         was re-selected on every poll cycle forever. It must now be marked terminal
-        exactly once, without being billed (no output means nothing to bill).
+        exactly once, without being billed: request_counts.completed == 0 proves the
+        missing output file means nothing to bill rather than a lagging output id
+        (#37713 keeps the lagging case eligible for the next cycle).
         """
         import base64
         from unittest.mock import patch
@@ -1073,6 +1075,7 @@ async def test_completed_without_output_file_marked_processed_without_billing(
         mock_response.status = completed_status
         mock_response.output_file_id = None
         mock_response.error_file_id = "file-error-123"
+        mock_response.request_counts = MagicMock(completed=0, failed=3, total=3)
         mock_response.model_dump_json.return_value = (
             f'{{"id":"batch-1","status":"{completed_status}"}}'
         )
@@ -1107,6 +1110,72 @@ async def test_completed_without_output_file_marked_processed_without_billing(
             mock_llm_router.get_deployment_credentials_with_provider.call_count == 0
         ), "a batch with no output file must not enter the cost-tracking path"
 
+    @pytest.mark.asyncio
+    @pytest.mark.parametrize(
+        "request_counts",
+        [MagicMock(completed=7, failed=0, total=7), None],
+        ids=["lagging_output_id", "unknown_counts"],
+    )
+    async def test_completed_with_lagging_output_file_left_for_next_cycle(
+        self,
+        check_batch_cost_instance,
+        mock_prisma_client,
+        mock_llm_router,
+        request_counts,
+    ):
+        """#37713 regression: a batch can report completed while its output_file_id is
+        still lagging behind at the provider. Retiring it in that window (or when the
+        request counts cannot prove there is nothing to bill) permanently loses the
+        spend record, so the poller must leave the row untouched and revisit it on the
+        next cycle once the output id has appeared.
+        """
+        import base64
+        from unittest.mock import patch
+
+        mock_prisma_client.db.litellm_managedobjecttable.update_many = AsyncMock(
+            return_value=1
+        )
+        mock_prisma_client.db.litellm_managedobjecttable.update = AsyncMock()
+        mock_prisma_client.db.litellm_usertable.find_unique = AsyncMock(
+            return_value=None
+        )
+
+        mock_job = MagicMock()
+        mock_job.id = "job-completed-lagging-output-1"
+        mock_job.unified_object_id = base64.urlsafe_b64encode(
+            b"litellm_proxy;model_id:model-123;llm_batch_id:batch-456"
+        ).decode()
+        mock_job.created_by = "user-1"
+
+        assert check_batch_cost_instance._has_batch_processed_column is True
+        mock_prisma_client.db.litellm_managedobjecttable.find_many = AsyncMock(
+            return_value=[mock_job]
+        )
+
+        mock_response = MagicMock()
+        mock_response.status = "completed"
+        mock_response.output_file_id = None
+        mock_response.error_file_id = None
+        mock_response.request_counts = request_counts
+
+        mock_llm_router.aretrieve_batch = AsyncMock(return_value=mock_response)
+        mock_llm_router.get_deployment_credentials_with_provider = MagicMock(
+            return_value={"api_key": "sk-test"}
+        )
+
+        with patch(
+            "litellm.files.main.afile_content",
+            new_callable=AsyncMock,
+        ) as mock_afile_content:
+            await check_batch_cost_instance.check_batch_cost()
+
+        assert (
+            mock_prisma_client.db.litellm_managedobjecttable.update.call_count == 0
+        ), "a completed batch whose output id is still lagging must stay eligible for the next poll"
+        assert (
+            mock_afile_content.await_count == 0
+        ), "a batch with no output file must not be billed"
+
     @pytest.mark.asyncio
     async def test_non_terminal_status_left_unprocessed(
         self, check_batch_cost_instance, mock_prisma_client, mock_llm_router
diff --git a/tests/test_litellm/proxy/openai_files_endpoint/test_files_common_utils.py b/tests/test_litellm/proxy/openai_files_endpoint/test_files_common_utils.py
index 2161e345b409..7f84407f8b39 100644
--- a/tests/test_litellm/proxy/openai_files_endpoint/test_files_common_utils.py
+++ b/tests/test_litellm/proxy/openai_files_endpoint/test_files_common_utils.py
@@ -428,3 +428,47 @@ def test_add_internal_model_credentials_survives_a_failing_deployment_lookup():
     add_internal_model_credentials(data=data, llm_router=router, model_id="deployment-gone")
 
     assert data == {"batch_id": "unified-batch-id"}
+
+
+from litellm.proxy.openai_files_endpoints.common_utils import (
+    _completed_batch_safe_to_retire,
+)
+
+
+def _completed_batch_for_retire(
+    output_file_id: str | None, completed: int | None = None
+) -> LiteLLMBatch:
+    kwargs = dict(
+        id="batch-1",
+        completion_window="24h",
+        created_at=1234567890,
+        endpoint="/v1/chat/completions",
+        input_file_id="file-in",
+        object="batch",
+        status="completed",
+        output_file_id=output_file_id,
+        error_file_id=None,
+    )
+    if completed is not None:
+        kwargs["request_counts"] = {"total": completed, "completed": completed, "failed": 0}
+    return LiteLLMBatch(**kwargs)
+
+
+class TestCompletedBatchSafeToRetire:
+    """A completed batch is only safe to retire from cost recovery once its output
+    file has arrived or the provider proves no successful lines (#37713)."""
+
+    def test_output_file_present_is_safe(self):
+        assert _completed_batch_safe_to_retire(_completed_batch_for_retire("file-out")) is True
+
+    def test_no_output_and_no_successful_lines_is_safe(self):
+        # Every request line errored -> nothing left to recover.
+        assert _completed_batch_safe_to_retire(_completed_batch_for_retire(None, completed=0)) is True
+
+    def test_no_output_but_successful_lines_is_not_safe(self):
+        # The bug: output_file_id is lagging; retiring here loses the spend record.
+        assert _completed_batch_safe_to_retire(_completed_batch_for_retire(None, completed=5)) is False
+
+    def test_no_output_and_unknown_counts_is_not_safe(self):
+        # Counts unknown -> stay eligible so the next poller pass revisits it.
+        assert _completed_batch_safe_to_retire(_completed_batch_for_retire(None)) is False