fix(cost_calc): update custom_llm_provider when base_model has different provider prefix (#22906)

* fix(cost_calc): update custom_llm_provider when base_model has different provider

When base_model carries a provider prefix that differs from the
deployment provider (e.g. base_model='gemini/gemini-2.0-flash' on an
anthropic/ deployment), the custom_llm_provider was not updated,
causing cost_per_token to build an invalid model key and return 0.

After _select_model_name_for_cost_calc resolves the model name from
base_model, extract the provider prefix and update custom_llm_provider
so the downstream cost lookup uses the correct provider.

Fixes #22257

Co-authored-by: Copilot <223556219+Copilot@users.noreply.github.com>

* fix(cost_calc): gate base_model provider override on custom_pricing and diff check

- Skip custom_llm_provider override when custom_pricing=True (base_model unused)
- Only override when extracted provider differs from current custom_llm_provider
- Add direct completion_cost unit test for cross-provider base_model
- Add same-provider no-regression test (e.g. openai/gpt-4o on openai deployment)

Co-authored-by: Copilot <223556219+Copilot@users.noreply.github.com>

* fix: guard hidden_params override when provider already overridden by base_model

- Add _provider_overridden flag to prevent hidden_params from undoing base_model fix
- Add direct unit test verifying hidden_params doesn't override extracted provider

Co-authored-by: Copilot <223556219+Copilot@users.noreply.github.com>

* fix: gate reasoning_auto_summary injection on supports_reasoning()

When reasoning_auto_summary is globally enabled, the reasoning param
was injected unconditionally for all models including non-reasoning
ones (e.g. gpt-4o-mini), causing OpenAI API errors. Now gated on
supports_reasoning(model, custom_llm_provider) check.

Co-authored-by: Copilot <223556219+Copilot@users.noreply.github.com>

* fix: revert unrelated transformation.py changes causing TypeError

The transformation.py changes pass model= and custom_llm_provider=
kwargs to _map_optional_params_to_responses_api_request() which only
accepts (self, optional_params, responses_api_request) — causing a
TypeError at runtime for every Responses API request.

Reverted to upstream version; cost_calculator.py fix is unaffected.

Co-authored-by: Copilot <223556219+Copilot@users.noreply.github.com>

---------

Co-authored-by: giulio-leone <6887247+giulio-leone@users.noreply.github.com>
Co-authored-by: Copilot <223556219+Copilot@users.noreply.github.com>

litellm/cost_calculator.py

@@ -1093,6 +1093,19 @@ def completion_cost(  # noqa: PLR0915
             router_model_id=router_model_id,
         )
 
+        # When base_model overrides model and carries its own provider prefix
+        # (e.g. base_model="gemini/gemini-2.0-flash" on an anthropic deployment),
+        # align custom_llm_provider so cost_per_token builds the correct key.
+        # Skip when custom_pricing is True (base_model is ignored in that path).
+        _provider_overridden = False
+        if base_model is not None and selected_model is not None and not custom_pricing:
+            _parts = selected_model.split("/", 1)
+            if len(_parts) > 1 and _parts[0] in LlmProvidersSet:
+                extracted = _parts[0]
+                if extracted != custom_llm_provider:
+                    custom_llm_provider = extracted
+                    _provider_overridden = True
+
         potential_model_names = [
             selected_model,
             _get_response_model(completion_response),
@@ -1174,9 +1187,10 @@ def completion_cost(  # noqa: PLR0915
 
                     hidden_params = getattr(completion_response, "_hidden_params", None)
                     if hidden_params is not None:
-                        custom_llm_provider = hidden_params.get(
-                            "custom_llm_provider", custom_llm_provider or None
-                        )
+                        if not _provider_overridden:
+                            custom_llm_provider = hidden_params.get(
+                                "custom_llm_provider", custom_llm_provider or None
+                            )
                         region_name = hidden_params.get("region_name", region_name)
 
                         # For Gemini/Vertex AI responses, trafficType is stored in

tests/local_testing/test_completion_cost.py

@@ -2945,3 +2945,77 @@ def test_batch_cost_calculator():
 
     cost = completion_cost(**args)
     assert cost > 0
+
+
+def test_cost_calculator_base_model_cross_provider():
+    """
+    When base_model has a different provider prefix than the deployment,
+    custom_llm_provider should be updated so cost_per_token builds the
+    correct model key.  Regression test for #22257.
+    """
+    resp = litellm.completion(
+        model="anthropic/my-custom-deployment",
+        messages=[{"role": "user", "content": "Hello"}],
+        base_model="gemini/gemini-2.0-flash",
+        mock_response="Hi there!",
+    )
+    assert resp._hidden_params["response_cost"] > 0
+
+
+def test_cost_calculator_base_model_cross_provider_direct():
+    """
+    Direct completion_cost unit test for cross-provider base_model override.
+    Verifies that completion_cost correctly routes to the base_model provider.
+    """
+    from litellm import ModelResponse, Usage
+
+    resp = ModelResponse(
+        id="chatcmpl-test",
+        model="gemini/gemini-2.0-flash",
+        usage=Usage(prompt_tokens=10, completion_tokens=20, total_tokens=30),
+    )
+    cost = completion_cost(
+        model="anthropic/my-custom-deployment",
+        completion_response=resp,
+        base_model="gemini/gemini-2.0-flash",
+        custom_llm_provider="anthropic",
+    )
+    assert cost > 0
+
+
+def test_cost_calculator_base_model_cross_provider_hidden_params_guard():
+    """
+    Verify that hidden_params.custom_llm_provider does not undo the
+    base_model provider override when the response carries a stale provider.
+    """
+    from litellm import ModelResponse, Usage
+
+    resp = ModelResponse(
+        id="chatcmpl-guard",
+        model="gemini/gemini-2.0-flash",
+        usage=Usage(prompt_tokens=10, completion_tokens=20, total_tokens=30),
+    )
+    # Simulate hidden_params carrying the original (wrong) provider
+    resp._hidden_params = {"custom_llm_provider": "anthropic"}
+
+    cost = completion_cost(
+        model="anthropic/my-custom-deployment",
+        completion_response=resp,
+        base_model="gemini/gemini-2.0-flash",
+        custom_llm_provider="anthropic",
+    )
+    assert cost > 0
+
+
+def test_cost_calculator_base_model_same_provider_no_regression():
+    """
+    When base_model has the same provider prefix as the deployment,
+    custom_llm_provider should remain unchanged (no-regression).
+    """
+    resp = litellm.completion(
+        model="openai/my-custom-deployment",
+        messages=[{"role": "user", "content": "Hello"}],
+        base_model="openai/gpt-4o",
+        mock_response="Hi there!",
+    )
+    assert resp._hidden_params["response_cost"] > 0