diff --git a/litellm/proxy/utils.py b/litellm/proxy/utils.py
index 874dca4238f..f891c85320a 100644
--- a/litellm/proxy/utils.py
+++ b/litellm/proxy/utils.py
@@ -7545,9 +7545,17 @@ def _group_token_limit(candidate_sets: tuple[tuple[ModelInfo, ...], ...], field:
     """The widest limit any deployment behind the listed name declares for ``field``.
 
     A model group is normally one model behind several interchangeable deployments, so
-    there is a single value to report. When a group genuinely mixes models, reporting the
-    widest window keeps the listing independent of config order and agreeing with
-    ``/model_group/info``, which aggregates the same way for the Admin UI.
+    there is a single value to report and the choice of aggregate does not arise.
+
+    When a group genuinely mixes models no single number is right, and the widest is the
+    deliberate pick over the narrowest for two reasons. It is what ``/model_group/info``
+    has long reported to the Admin UI, so the two surfaces agree; disagreeing is the very
+    complaint this resolution path exists to fix. And of the two ways to be wrong,
+    under-advertising is worse: a client that trusts a narrowed window silently refuses
+    prompts the group would have served, while an over-long prompt that reaches a smaller
+    deployment comes back as a legible context-length error -- and does not reach one at
+    all when ``enable_pre_call_checks`` is set, which filters deployments the prompt does
+    not fit.
     """
     limits: Final = tuple(
         limit for limit in (_first_token_limit(candidates, field) for candidates in candidate_sets) if limit is not None
diff --git a/tests/test_litellm/test_router.py b/tests/test_litellm/test_router.py
index 588fd98abd1..eba9a85feb6 100644
--- a/tests/test_litellm/test_router.py
+++ b/tests/test_litellm/test_router.py
@@ -7350,6 +7350,20 @@ def test_get_model_listing_info_carries_configured_limits():
     assert (info.max_input_tokens, info.max_output_tokens) == (32000, 8000)
 
 
+def test_widest_configured_limit_ignores_absent_and_malformed_values():
+    model_infos = (
+        {"max_input_tokens": 32000},
+        {},
+        {"max_input_tokens": "not-a-number"},
+        {"max_input_tokens": "128000"},
+        {"max_output_tokens": 4096},
+    )
+
+    assert litellm.Router._widest_configured_limit(model_infos, "max_input_tokens") == 128000
+    assert litellm.Router._widest_configured_limit(model_infos, "max_output_tokens") == 4096
+    assert litellm.Router._widest_configured_limit((), "max_input_tokens") is None
+
+
 def test_get_model_listing_info_dedupes_interchangeable_deployments():
     """The ordinary group is N deployments of one model, so it yields exactly one key."""
     router = litellm.Router(