diff --git a/litellm/integrations/newrelic/newrelic_metrics.py b/litellm/integrations/newrelic/newrelic_metrics.py
index 5b938c847d5..7db26eff0fc 100644
--- a/litellm/integrations/newrelic/newrelic_metrics.py
+++ b/litellm/integrations/newrelic/newrelic_metrics.py
@@ -61,6 +61,10 @@ from litellm.types.integrations.newrelic import (
 )
 from litellm.types.utils import StandardLoggingPayload
 
+# 408 (request timeout) and 429 (rate limit) are transient client errors the
+# Metric API expects a retry on, unlike 400/403 which a retry would only repeat.
+_RETRYABLE_CLIENT_STATUSES: Final = frozenset({408, 429})
+
 
 def resolve_newrelic_metric_endpoint(newrelic_region: str | None) -> str:
     if not newrelic_region:
@@ -346,7 +350,7 @@ class NewRelicMetricsLogger(CustomBatchLogger):
         if 200 <= status < 300:
             return True
 
-        if 400 <= status < 500:
+        if 400 <= status < 500 and status not in _RETRYABLE_CLIENT_STATUSES:
             verbose_logger.warning(
                 "New Relic Metrics: %s from Metric API%s, dropping %s records.",
                 status,
diff --git a/tests/test_litellm/integrations/newrelic/test_newrelic_metrics.py b/tests/test_litellm/integrations/newrelic/test_newrelic_metrics.py
index e029135d2a2..f28e030f77f 100644
--- a/tests/test_litellm/integrations/newrelic/test_newrelic_metrics.py
+++ b/tests/test_litellm/integrations/newrelic/test_newrelic_metrics.py
@@ -770,6 +770,20 @@ async def test_raised_500_is_requeued():
     assert logger.log_queue == [record], "a transient 5xx must requeue"
 
 
+@pytest.mark.asyncio
+@pytest.mark.parametrize("status", [429, 408])
+async def test_transient_4xx_is_requeued_not_dropped(status):
+    """The Metric API returns 429 when it throttles (and 408 on a request
+    timeout); both are transient and expect a retry, so the batch must be
+    requeued rather than permanently dropped like a 400/403."""
+    logger = _make_logger()
+    record = _record()
+    logger.log_queue.append(record)
+    logger.async_client.post = _raises(status)
+    await logger.async_send_batch()
+    assert logger.log_queue == [record], f"a transient {status} must requeue, not drop"
+
+
 @pytest.mark.asyncio
 @pytest.mark.parametrize("status", [200, 201, 204])
 async def test_any_2xx_is_treated_as_delivered_not_requeued(status):