diff --git a/openagent_eval/metrics/performance/latency.py b/openagent_eval/metrics/performance/latency.py
index face4c1..58c7c7f 100644
--- a/openagent_eval/metrics/performance/latency.py
+++ b/openagent_eval/metrics/performance/latency.py
@@ -40,9 +40,16 @@ def evaluate(self, **kwargs: Any) -> MetricResult:
         Returns:
             MetricResult with latency score.
         """
-        latency_ms = kwargs.get("latency_ms", 0.0)
+        latency_ms = kwargs.get("latency_ms")
         stage = kwargs.get("stage", "unknown")
 
+        if latency_ms is None:
+            return MetricResult(
+                score=0.0,
+                reason="No latency measurement available",
+                metadata={"stage": stage, "latency_ms": None},
+            )
+
         if latency_ms < 0:
             return MetricResult(
                 score=0.0,
diff --git a/tests/unit/test_metrics/test_performance_cost.py b/tests/unit/test_metrics/test_performance_cost.py
index 43181a6..1f9c329 100644
--- a/tests/unit/test_metrics/test_performance_cost.py
+++ b/tests/unit/test_metrics/test_performance_cost.py
@@ -56,6 +56,17 @@ def test_metadata(self):
         assert result.metadata["stage"] == "embedding"
         assert result.metadata["latency_ms"] == 100.0
 
+    def test_none_latency_does_not_raise(self):
+        """None latency (e.g. no LLM configured, or generation failure) is handled gracefully."""
+        result = self.metric.evaluate(latency_ms=None, stage="llm")
+        assert result.score == 0.0
+        assert result.metadata["latency_ms"] is None
+
+    def test_missing_latency_kwarg_does_not_raise(self):
+        """Omitting latency_ms entirely behaves the same as passing None."""
+        result = self.metric.evaluate(stage="llm")
+        assert result.score == 0.0
+
 
 class TestTokenCountMetric:
     """Tests for TokenCountMetric."""