fix(cost): apply vertex regional uplift to cost breakdown line items

Co-Authored-By: Devin AI <158243242+devin-ai-integration[bot]@users.noreply.github.com>

litellm/cost_calculator.py

@@ -1661,6 +1661,7 @@ def completion_cost(
                             usage=cost_per_token_usage_object,
                             service_tier=service_tier,
                             data_residency=data_residency,
+                            vertex_location=vertex_location,
                         )
                         _reasoning_cost = _token_type_breakdown.reasoning_cost
                         _cache_read_cost = _token_type_breakdown.cache_read_cost

litellm/litellm_core_utils/llm_cost_calc/utils.py

@@ -694,6 +694,23 @@ def _get_regional_uplift_multiplier(model_info: ModelInfo, data_residency: str |
         return 1.0
 
 
+GLOBAL_VERTEX_LOCATION: Final = "global"
+
+
+def get_vertex_regional_endpoint_uplift(model_info: ModelInfo, vertex_location: str | None) -> float:
+    """
+    Vertex bills a flat premium (currently +10%) on every token type when a request is served
+    from a regional or multi-region endpoint instead of the global one, so the location the
+    request was routed to decides the rate, not just the model.
+    """
+    if vertex_location is None or vertex_location.lower() == GLOBAL_VERTEX_LOCATION:
+        return 1.0
+    multiplier: Final = model_info.get("regional_endpoint_uplift_multiplier")
+    if multiplier is None:
+        return 1.0
+    return float(multiplier)
+
+
 def _resolve_reasoning_token_cost(
     model_info: ModelInfo,
     service_tier: str | None,
@@ -903,6 +920,7 @@ def get_token_type_cost_breakdown(
     usage: Usage,
     service_tier: str | None = None,
     data_residency: str | None = None,
+    vertex_location: str | None = None,
 ) -> TokenTypeCostBreakdown:
     """
     Provider-agnostic cost of reasoning and cache tokens, derived from the usage
@@ -975,7 +993,9 @@ def get_token_type_cost_breakdown(
 
     # Apply the same flat regional-processing uplift the totals get, so per-type
     # costs stay reconciled with input_cost/output_cost for regionalized OpenAI hosts.
-    uplift: Final = _get_regional_uplift_multiplier(model_info, data_residency)
+    uplift: Final = _get_regional_uplift_multiplier(model_info, data_residency) * get_vertex_regional_endpoint_uplift(
+        model_info, vertex_location
+    )
     if uplift != 1.0:
         reasoning_cost *= uplift
         cache_read_cost *= uplift

litellm/llms/vertex_ai/cost_calculator.py

@@ -7,6 +7,7 @@ from litellm import verbose_logger
 from litellm.litellm_core_utils.llm_cost_calc.utils import (
     _is_above_128k,
     generic_cost_per_token,
+    get_vertex_regional_endpoint_uplift,
 )
 from litellm.types.utils import ModelInfo, Usage
 
@@ -26,22 +27,6 @@ Google AI Studio -> token based pricing
 
 models_without_dynamic_pricing: Final = ["gemini-1.0-pro", "gemini-pro", "gemini-2"]
 
-GLOBAL_VERTEX_LOCATION: Final = "global"
-
-
-def _regional_endpoint_uplift(model_info: ModelInfo, vertex_location: str | None) -> float:
-    """
-    Vertex bills a flat premium (currently +10%) on every token type when a request is served
-    from a regional or multi-region endpoint instead of the global one, so the location the
-    request was routed to decides the rate, not just the model.
-    """
-    if vertex_location is None or vertex_location.lower() == GLOBAL_VERTEX_LOCATION:
-        return 1.0
-    multiplier: Final = model_info.get("regional_endpoint_uplift_multiplier")
-    if multiplier is None:
-        return 1.0
-    return float(multiplier)
-
 
 def cost_router(
     model: str,
@@ -253,5 +238,5 @@ def cost_per_token(
             service_tier=service_tier,
         )
 
-    uplift: Final = _regional_endpoint_uplift(model_info=model_info, vertex_location=vertex_location)
+    uplift: Final = get_vertex_regional_endpoint_uplift(model_info, vertex_location)
     return prompt_cost * uplift, completion_cost * uplift

tests/test_litellm/litellm_core_utils/llm_cost_calc/test_llm_cost_calc_utils.py

@@ -2207,6 +2207,31 @@ def test_token_type_cost_breakdown_matches_real_gemini_numbers():
     assert breakdown.cache_creation_cost == 0.0
 
 
+@pytest.mark.parametrize("vertex_location, uplift", [("global", 1.0), ("us-east5", 1.1)])
+def test_token_type_cost_breakdown_applies_vertex_regional_uplift(vertex_location, uplift):
+    """Component costs must carry the same regional uplift as the total, or spend logs stop adding up."""
+    os.environ["LITELLM_LOCAL_MODEL_COST_MAP"] = "True"
+    litellm.model_cost = litellm.get_model_cost_map(url="")
+
+    usage = Usage(
+        prompt_tokens=2500,
+        completion_tokens=200,
+        total_tokens=2700,
+        cache_creation_input_tokens=500,
+        cache_read_input_tokens=2000,
+    )
+
+    breakdown = get_token_type_cost_breakdown(
+        model="claude-sonnet-4-6",
+        custom_llm_provider="vertex_ai",
+        usage=usage,
+        vertex_location=vertex_location,
+    )
+
+    assert breakdown.cache_read_cost == pytest.approx(2000 * 3e-07 * uplift)
+    assert breakdown.cache_creation_cost == pytest.approx(500 * 3.75e-06 * uplift)
+
+
 def test_token_type_cost_breakdown_xai_at_exactly_200k_uses_higher_tier_rates():
     os.environ["LITELLM_LOCAL_MODEL_COST_MAP"] = "True"
     litellm.model_cost = litellm.get_model_cost_map(url="")