diff --git a/tests/test_litellm/integrations/test_prometheus_v3_remaining_headers.py b/tests/test_litellm/integrations/test_prometheus_v3_remaining_headers.py
new file mode 100644
index 00000000000..6cd10a911b0
--- /dev/null
+++ b/tests/test_litellm/integrations/test_prometheus_v3_remaining_headers.py
@@ -0,0 +1,97 @@
+"""
+Regression test for: Prometheus per-key remaining gauges show 9e18 (sys.maxsize)
+when parallel_request_limiter_v3 is in use.
+
+The v3 limiter publishes remaining counters on the response under
+hidden_params.additional_headers as
+    "x-ratelimit-model_per_key-remaining-{requests,tokens}"
+rather than into request metadata under
+    "litellm-key-remaining-{requests,tokens}-{model_group}".
+
+prometheus.py must read both, otherwise DataDog / Prometheus display
+sys.maxsize (~9.22e18) instead of the real remaining values.
+"""
+import os
+import sys
+
+import pytest
+from prometheus_client import REGISTRY
+
+sys.path.insert(0, os.path.abspath("../../.."))
+
+from litellm.integrations.prometheus import PrometheusLogger
+
+
+@pytest.fixture(scope="function")
+def prometheus_logger():
+    for c in list(REGISTRY._collector_to_names.keys()):
+        try:
+            REGISTRY.unregister(c)
+        except Exception:
+            pass
+    return PrometheusLogger()
+
+
+class TestVirtualKeyRateLimitMetricsV3Headers:
+    def _call(self, logger, *, metadata, additional_headers):
+        kwargs = {
+            "litellm_params": {
+                "metadata": {"model_group": "gpt-4o"},
+                "model_group": "gpt-4o",
+            },
+            "model": "gpt-4o",
+        }
+        slp = {
+            "hidden_params": {"additional_headers": additional_headers}
+        }
+        logger._set_virtual_key_rate_limit_metrics(
+            user_api_key="hashed-key",
+            user_api_key_alias="alias",
+            kwargs=kwargs,
+            metadata=metadata,
+            model_id="m1",
+            standard_logging_payload=slp,
+        )
+        labels = ("hashed-key", "alias", "gpt-4o", "m1")
+        req = logger.litellm_remaining_api_key_requests_for_model.labels(*labels)._value.get()
+        tok = logger.litellm_remaining_api_key_tokens_for_model.labels(*labels)._value.get()
+        return req, tok
+
+    def test_falls_back_to_v3_additional_headers(self, prometheus_logger):
+        """When metadata lacks legacy keys but v3 headers are set, use them."""
+        metadata = {"model_group": "gpt-4o"}  # no litellm-key-remaining-*
+        headers = {
+            "x-ratelimit-model_per_key-remaining-tokens": 1234,
+            "x-ratelimit-model_per_key-remaining-requests": 7,
+        }
+        req, tok = self._call(
+            prometheus_logger, metadata=metadata, additional_headers=headers
+        )
+        assert req == 7
+        assert tok == 1234
+        assert req < 1e18 and tok < 1e18, "must not be sys.maxsize"
+
+    def test_metadata_keys_take_precedence(self, prometheus_logger):
+        """If legacy metadata keys are present, prefer them (back-compat)."""
+        metadata = {
+            "model_group": "gpt-4o",
+            "litellm-key-remaining-requests-gpt-4o": 11,
+            "litellm-key-remaining-tokens-gpt-4o": 2222,
+        }
+        headers = {
+            "x-ratelimit-model_per_key-remaining-tokens": 99,
+            "x-ratelimit-model_per_key-remaining-requests": 99,
+        }
+        req, tok = self._call(
+            prometheus_logger, metadata=metadata, additional_headers=headers
+        )
+        assert req == 11
+        assert tok == 2222
+
+    def test_no_signals_falls_back_to_maxsize(self, prometheus_logger):
+        """Preserve old behaviour: when nothing is set, default to sys.maxsize."""
+        req, tok = self._call(
+            prometheus_logger, metadata={"model_group": "gpt-4o"}, additional_headers={}
+        )
+        assert req == float(sys.maxsize)
+        assert tok == float(sys.maxsize)