diff --git a/litellm/integrations/prometheus.py b/litellm/integrations/prometheus.py
index 2af5faf0757..e513d2c655e 100644
--- a/litellm/integrations/prometheus.py
+++ b/litellm/integrations/prometheus.py
@@ -2617,9 +2617,9 @@ class PrometheusLogger(CustomLogger):
         # With no deployment behind the call, the name is whatever the caller sent (for example a model the
         # key may not use). Bound it like ``requested_model`` so unknown names share one series.
         model: Final = (
-            _bounded_requested_model_label(requested_name, router_originated=True)
-            if not deployment_model_id and isinstance(requested_name, str)
-            else requested_name
+            requested_name
+            if deployment_model_id
+            else _bounded_requested_model_label(requested_name, router_originated=True)  # pyright: ignore[reportUnknownArgumentType]  # caller-supplied value, the helper type-checks it
         )
 
         litellm_params: Final = kwargs.get("litellm_params", {}) or {}
diff --git a/tests/unit/integrations/test_prometheus_requested_model_cardinality.py b/tests/unit/integrations/test_prometheus_requested_model_cardinality.py
index de08ff556ec..f4f0d3607e4 100644
--- a/tests/unit/integrations/test_prometheus_requested_model_cardinality.py
+++ b/tests/unit/integrations/test_prometheus_requested_model_cardinality.py
@@ -360,3 +360,14 @@ async def test_sdk_failures_without_proxy_router_keep_their_model_label_on_llm_a
         await _fire_llm_api_failure(logger, "gpt-4o-mini")
 
     assert _model_label_values(logger.litellm_llm_api_failed_requests_metric) == {"gpt-4o-mini"}
+
+
+@pytest.mark.asyncio
+@pytest.mark.parametrize("model", [123, 4.5, ["gpt-4o-mini"], {"name": "gpt-4o-mini"}])
+async def test_non_string_denied_models_collapse_to_other_on_llm_api_failed_requests_metric(router, model: object):
+    logger = PrometheusLogger()
+
+    with patch("litellm.proxy.proxy_server.llm_router", router, create=True):  # test-quality-ok: production reads proxy_server.llm_router lazily, no injection seam
+        await _fire_llm_api_failure(logger, model)  # type: ignore[arg-type]
+
+    assert _model_label_values(logger.litellm_llm_api_failed_requests_metric) == {UNRECOGNIZED_REQUESTED_MODEL_LABEL}