diff --git a/litellm/proxy/common_utils/periodic_reload_schedule.py b/litellm/proxy/common_utils/periodic_reload_schedule.py
index 26ae6bbc4a7..1712376fb14 100644
--- a/litellm/proxy/common_utils/periodic_reload_schedule.py
+++ b/litellm/proxy/common_utils/periodic_reload_schedule.py
@@ -80,8 +80,17 @@ class _IntervalConfig(BaseModel):
     interval_hours: int | None = None
 
 
+def to_db_precision(value: datetime) -> datetime:
+    """
+    Drop sub-millisecond digits, matching the TIMESTAMP(3) columns these values round-trip
+    through. Comparing a microsecond-precision stamp against its own truncated copy would
+    read as newer and skip the request it recorded
+    """
+    return value.replace(microsecond=(value.microsecond // 1000) * 1000)
+
+
 def utc_now() -> datetime:
-    return datetime.now(timezone.utc)
+    return to_db_precision(datetime.now(timezone.utc))
 
 
 def _parse_interval_hours(param_value: object) -> int | None:
diff --git a/litellm/proxy/proxy_server.py b/litellm/proxy/proxy_server.py
index a28f39c7a86..4f75763061e 100644
--- a/litellm/proxy/proxy_server.py
+++ b/litellm/proxy/proxy_server.py
@@ -333,6 +333,7 @@ from litellm.proxy.common_utils.periodic_reload_schedule import (
     record_manual_reload,
     record_reload_run,
     reload_schedule_status,
+    to_db_precision,
     utc_now,
     write_reload_interval,
 )
@@ -3862,7 +3863,10 @@ class ProxyConfig:
             get_model_cost_map_loaded_at,
         )
 
-        self.model_cost_map_loaded_at: datetime = get_model_cost_map_loaded_at() or utc_now()
+        cost_map_loaded_at = get_model_cost_map_loaded_at()
+        self.model_cost_map_loaded_at: datetime = (
+            to_db_precision(cost_map_loaded_at) if cost_map_loaded_at is not None else utc_now()
+        )
 
     def is_yaml(self, config_file_path: str) -> bool:
         if not os.path.isfile(config_file_path):
diff --git a/tests/test_litellm/proxy/common_utils/test_periodic_reload_schedule.py b/tests/test_litellm/proxy/common_utils/test_periodic_reload_schedule.py
index ab28776ee29..e282059eac2 100644
--- a/tests/test_litellm/proxy/common_utils/test_periodic_reload_schedule.py
+++ b/tests/test_litellm/proxy/common_utils/test_periodic_reload_schedule.py
@@ -13,6 +13,8 @@ from litellm.proxy.common_utils.periodic_reload_schedule import (
     record_manual_reload,
     record_reload_run,
     reload_schedule_status,
+    to_db_precision,
+    utc_now,
     write_reload_interval,
 )
 
@@ -134,6 +136,32 @@ def test_pod_reload_is_due(schedule, pod_data_loaded_at, expected):
     )
 
 
+def test_stamps_match_the_precision_they_are_stored_at():
+    """Postgres stores TIMESTAMP(3), so a microsecond-precision pod clock would read as
+    newer than its own persisted copy and skip the request it just recorded"""
+    assert utc_now().microsecond % 1000 == 0
+    assert to_db_precision(datetime(2024, 1, 1, 6, 0, 0, 123456, tzinfo=timezone.utc)) == datetime(
+        2024, 1, 1, 6, 0, 0, 123000, tzinfo=timezone.utc
+    )
+
+
+def test_request_in_the_same_millisecond_is_not_lost_to_truncation():
+    """A pod whose data loaded microseconds before a request must still see it: both sides
+    are floored to the stored resolution, so the request never reads as older than it is"""
+    loaded_at = to_db_precision(datetime(2024, 1, 1, 6, 0, 0, 123400, tzinfo=timezone.utc))
+    requested_at = datetime(2024, 1, 1, 6, 0, 0, 124000, tzinfo=timezone.utc)
+
+    assert (
+        pod_reload_is_due(
+            schedule=ReloadSchedule(reload_requested_at=requested_at),
+            pod_data_loaded_at=loaded_at,
+            current_time=NOW,
+            description="test",
+        )
+        is True
+    )
+
+
 def test_manual_request_reaches_every_pod_with_older_data():
     """Nothing clears reload_requested_at, so each pod reloads exactly once per request:
     due while its data predates the request, not due once refreshed or booted after it"""