From 6c313e9b5e5d088c0d3f5fd124a1d347144266e3 Mon Sep 17 00:00:00 2001
From: mateo-berri <277851410+mateo-berri@users.noreply.github.com>
Date: Mon, 17 Aug 2026 16:48:43 +0000
Subject: [PATCH] feat(cost-calc): infer data_residency from response
 hidden_params
MIME-Version: 1.0
Content-Type: text/plain; charset=UTF-8
Content-Transfer-Encoding: 8bit

completion_cost(completion_response=resp) is a user-facing entry point
that reruns the cost calculation from the response object alone — it
does not have access to the logging object's litellm_params. Without
this, users calling completion_cost() after a completion would see the
base cost, while the internal cost calculation (via _response_cost_calculator
in litellm_logging.py) would apply the uplift and store it on
_hidden_params["response_cost"]. Symmetry: also read api_base from
hidden_params and infer data_residency there.

https://claude.ai/code/session_012ebH44s7ohYxjoix5CXzTW
---
 litellm/cost_calculator.py                    | 10 +++++
 .../llm_cost_calc/test_llm_cost_calc_utils.py | 42 +++++++++++++++++++
 2 files changed, 52 insertions(+)

diff --git a/litellm/cost_calculator.py b/litellm/cost_calculator.py
index 2e7cbf4d039..45d1ec131c2 100644
--- a/litellm/cost_calculator.py
+++ b/litellm/cost_calculator.py
@@ -15,6 +15,7 @@ from litellm.constants import (
     DEFAULT_MAX_LRU_CACHE_SIZE,
     DEFAULT_REPLICATE_GPU_PRICE_PER_SECOND,
 )
+from litellm.litellm_core_utils.data_residency import infer_openai_data_residency
 from litellm.litellm_core_utils.llm_cost_calc.tool_call_cost_tracking import (
     StandardBuiltInToolCostTracking,
 )
@@ -1302,6 +1303,15 @@ def completion_cost(  # noqa: PLR0915
                         )
                         region_name = hidden_params.get("region_name", region_name)
 
+                        # For OpenAI, infer data_residency from api_base if it
+                        # wasn't passed in explicitly. Keeps external
+                        # completion_cost(response) in sync with the internal
+                        # cost calculation (which reads it from litellm_params).
+                        if data_residency is None and custom_llm_provider == "openai":
+                            data_residency = infer_openai_data_residency(
+                                hidden_params.get("api_base")
+                            )
+
                         # For Gemini/Vertex AI responses, trafficType is stored in
                         # provider_specific_fields.  Map it to the service_tier used
                         # by the cost key lookup (_priority / _flex suffixes) so that
diff --git a/tests/test_litellm/litellm_core_utils/llm_cost_calc/test_llm_cost_calc_utils.py b/tests/test_litellm/litellm_core_utils/llm_cost_calc/test_llm_cost_calc_utils.py
index dc6607e7009..667baf65ed3 100644
--- a/tests/test_litellm/litellm_core_utils/llm_cost_calc/test_llm_cost_calc_utils.py
+++ b/tests/test_litellm/litellm_core_utils/llm_cost_calc/test_llm_cost_calc_utils.py
@@ -1530,3 +1530,45 @@ def test_data_residency_composes_with_service_tier(_local_model_cost_map):
 
     assert priority_base_total > 0
     assert priority_eu_total == pytest.approx(priority_base_total * 1.10, rel=1e-9)
+
+
+def test_completion_cost_infers_data_residency_from_hidden_params(
+    _local_model_cost_map,
+):
+    """External completion_cost(response) infers data_residency from
+    the api_base recorded on hidden_params, so it matches the internal cost
+    calculation from the completion flow."""
+    from litellm import ModelResponse, completion_cost
+    from litellm.types.utils import Choices, Message
+    from litellm.types.utils import Usage as U
+
+    def _resp(api_base):
+        r = ModelResponse(
+            id="test",
+            choices=[
+                Choices(
+                    finish_reason="stop",
+                    index=0,
+                    message=Message(content="ok", role="assistant"),
+                )
+            ],
+            model="gpt-5",
+            object="chat.completion",
+            created=0,
+            usage=U(prompt_tokens=1000, completion_tokens=500, total_tokens=1500),
+        )
+        r._hidden_params = {
+            "custom_llm_provider": "openai",
+            "api_base": api_base,
+        }
+        return r
+
+    global_cost = completion_cost(
+        completion_response=_resp("https://api.openai.com/v1")
+    )
+    eu_cost = completion_cost(completion_response=_resp("https://eu.api.openai.com/v1"))
+    us_cost = completion_cost(completion_response=_resp("https://us.api.openai.com/v1"))
+
+    assert global_cost > 0
+    assert eu_cost == pytest.approx(global_cost * 1.10, rel=1e-9)
+    assert us_cost == pytest.approx(global_cost * 1.10, rel=1e-9)