fix(cost): stop non-string service_tier from silently dropping cost tracking (#30690)

completion_cost read service_tier straight from the request optional_params
and called service_tier.lower() on it, so a non-string value (dict/int/list,
reachable via allowed_openai_params/drop_params) raised AttributeError.
_response_cost_calculator swallowed that and returned response_cost=None, so
the request's cost was silently lost.

The isinstance guard alone is not enough: a surviving dict would crash again
downstream in _get_service_tier_cost_key, which also calls .lower(). A
request-level service_tier is only meaningful for pricing when it is a concrete
billable tier string, so coerce any non-string value to None and defer to the
tier the provider reports on the response usage, the same way "auto" already
does.

Adds a regression test driving a dict service_tier through completion_cost; it
raises AttributeError before the fix and prices at the served tier after.

(cherry picked from commit 43dadc5138)

litellm/cost_calculator.py

@@ -1227,10 +1227,14 @@ def completion_cost(  # noqa: PLR0915
         if service_tier is None and optional_params is not None:
             service_tier = optional_params.get("service_tier")
 
-        # "auto" is a routing preference, not a billable tier: the provider picks
-        # the tier and reports the one actually served on the response/usage, so
-        # defer to that instead of pricing the request-level "auto" as standard
-        if service_tier is not None and service_tier.lower() == ServiceTier.AUTO.value:
+        # A request-level service_tier only prices the request when it is a
+        # concrete billable tier string. "auto" is a routing preference and any
+        # non-string value is not a billable tier, so defer to the tier the
+        # provider reports on the response/usage instead of crashing or mispricing
+        if (
+            not isinstance(service_tier, str)
+            or service_tier.lower() == ServiceTier.AUTO.value
+        ):
             service_tier = None
 
         # Extract service_tier from completion_response if not provided

tests/test_litellm/test_cost_calculator.py

@@ -2054,6 +2054,58 @@ def test_completion_cost_anthropic_auto_tier_uses_served_priority_rate():
     assert cost == pytest.approx(expected_priority)
 
 
+def test_completion_cost_non_string_service_tier_defers_to_served_tier():
+    """
+    Regression: a non-string request-level ``service_tier`` (reachable via
+    ``allowed_openai_params``/``drop_params``) must not crash cost tracking.
+
+    Before the fix, ``completion_cost`` called ``service_tier.lower()`` on the
+    request-level value, so a dict raised ``AttributeError``. ``_response_cost_calculator``
+    swallowed it and reported ``response_cost=None``, silently dropping the cost.
+    The non-string preference must be ignored so pricing defers to the tier the
+    provider actually served on the response usage.
+    """
+    from litellm import completion_cost
+    from litellm.llms.anthropic.chat.transformation import AnthropicConfig
+
+    os.environ["LITELLM_LOCAL_MODEL_COST_MAP"] = "True"
+    litellm.model_cost = litellm.get_model_cost_map(url="")
+
+    model = "claude-test-non-string-tier-cost-model"
+    litellm.register_model(
+        model_cost={
+            model: {
+                "input_cost_per_token": 3e-6,
+                "output_cost_per_token": 15e-6,
+                "input_cost_per_token_priority": 6e-6,
+                "output_cost_per_token_priority": 30e-6,
+                "litellm_provider": "anthropic",
+                "max_tokens": 8192,
+            }
+        }
+    )
+
+    usage = AnthropicConfig().calculate_usage(
+        usage_object={
+            "input_tokens": 1000,
+            "output_tokens": 500,
+            "service_tier": "priority",
+        },
+        reasoning_content=None,
+    )
+    response = ModelResponse(usage=usage, model=model)
+
+    cost = completion_cost(
+        completion_response=response,
+        model=model,
+        custom_llm_provider="anthropic",
+        optional_params={"service_tier": {"name": "auto"}},
+    )
+
+    expected_priority = 1000 * 6e-6 + 500 * 30e-6
+    assert cost == pytest.approx(expected_priority)
+
+
 def test_anthropic_cost_per_token_prices_cache_at_served_tier_with_multiplier():
     """
     Regression for the cache/tier interaction in the Anthropic geo/speed path.