fix(router): ignore the azure gpt5_series/ selector when resolving model info

An Azure deployment such as azure/gpt5_series/gpt-5.6-sol with no base_model looked up azure/gpt5_series/gpt-5.6-sol in the cost map, found only the zeroed stub the Router registers for deployment names, and so got no context window, no pricing and no pre-call context check. The fallback now strips the selector before the cost-map lookup, and the request itself still keeps it.

litellm/router_python_core.py

@@ -223,6 +223,7 @@ from litellm.litellm_core_utils.token_counter import (
 )
 from litellm.litellm_core_utils.tool_fields import TOOL_FIELDS, drop_policy
 from litellm.llms.anthropic.common_utils import AnthropicModelInfo, anthropic_error_frame_exception
+from litellm.llms.azure.chat.gpt_reasoning_transformation import AzureOpenAIGPT5Config
 from litellm.llms.base_llm.passthrough.transformation import replace_path_segment
 from litellm.llms.base_llm.vector_store.transformation import (
     RouterVectorStoreEmbeddingExecutor,
@@ -12885,13 +12886,17 @@ class PythonCore(DeprecatedRouterMembers):
             # Router init auto-registers every deployment name into
             # litellm.model_cost as a zeroed stub, so membership alone can't
             # tell a resolvable name apart; require usable limits/costs.
-            _azure_fallback_key = _model if _model.startswith("azure/") else f"azure/{_model}"
+            _metadata_model: Final = _model.removeprefix(AzureOpenAIGPT5Config.GPT5_SERIES_ROUTE)
+            _azure_fallback_key: Final = (
+                _metadata_model if _metadata_model.startswith("azure/") else f"azure/{_metadata_model}"
+            )
             _fallback_entry = self._host.dynamic().model_cost.get(_azure_fallback_key)
             _fallback_resolves = isinstance(_fallback_entry, Mapping) and (
                 (_fallback_entry.get("max_input_tokens") or 0) > 0
                 or (_fallback_entry.get("input_cost_per_token") or 0) > 0
             )
             if _fallback_resolves:
+                model = _metadata_model
                 verbose_router_logger.debug(
                     "Azure deployment '%s' has no base_model set; using '%s' from the model cost map for max tokens, cost tracking, etc.",
                     _model,

tests/unit/test_router/test_router.py

@@ -17196,6 +17196,36 @@ class TestAzureBaseModelFallbackLogging:
         )
         assert model_info["max_input_tokens"] == litellm.model_cost["azure/gpt-4o-mini"]["max_input_tokens"]
 
+    def test_gpt5_series_selector_resolves_the_underlying_model(self):
+        router = self._router_with_azure_deployment("azure/gpt5_series/gpt-5.6-sol")
+
+        model_info = router.get_router_model_info(
+            deployment=None, received_model_name="my-group", id="azure-base-model-test-id"
+        )
+
+        assert model_info["max_input_tokens"] == litellm.model_cost["azure/gpt-5.6-sol"]["max_input_tokens"]
+        assert model_info["input_cost_per_token"] == litellm.model_cost["azure/gpt-5.6-sol"]["input_cost_per_token"]
+
+    def test_gpt5_series_selector_cannot_bypass_the_context_window_check(self):
+        router = self._router_with_azure_deployment("azure/gpt5_series/gpt-5.6-sol")
+        limit = litellm.model_cost["azure/gpt-5.6-sol"]["max_input_tokens"]
+        messages = [{"role": "user", "content": "hi"}]
+
+        with patch.object(litellm, "token_counter", return_value=limit):
+            router._pre_call_checks(model="my-group", healthy_deployments=router.model_list, messages=messages)
+        with patch.object(litellm, "token_counter", return_value=limit + 1):
+            with pytest.raises(litellm.ContextWindowExceededError):
+                router._pre_call_checks(model="my-group", healthy_deployments=router.model_list, messages=messages)
+
+    def test_gpt5_series_selector_does_not_give_an_unknown_deployment_any_limits(self):
+        router = self._router_with_azure_deployment("azure/gpt5_series/my-custom-deployment-name")
+
+        model_info = router.get_router_model_info(
+            deployment=None, received_model_name="my-group", id="azure-base-model-test-id"
+        )
+
+        assert model_info.get("max_input_tokens") is None
+
 def test_model_group_info_intersects_supported_reasoning_efforts():
     router = litellm.Router(
         model_list=[