fix(cost_calculator): respect custom pricing for known providers like anthropic

When custom_llm_provider is set to a known provider (e.g., "anthropic")
and the user configures explicit zero costs (input_cost_per_token: 0.0),
cost_per_token() dispatched directly to the provider-specific function
(anthropic_cost_per_token) before checking custom pricing entries. This
caused the zero-cost override to be ignored and real provider pricing to
be applied instead.

- Add custom_pricing parameter to cost_per_token() and pass it from
  completion_cost()
- When custom_pricing=True and model is in model_cost, bypass the
  provider dispatch chain and use generic_cost_per_token() directly
- Fix zero-cost check in else branch: use `is not None` instead of
  `or 0.0 > 0` which collapsed 0.0 to False

Fixes #25204

Co-Authored-By: Claude Opus 4.6 <noreply@anthropic.com>

litellm/cost_calculator.py

@@ -283,6 +283,7 @@ def cost_per_token(  # noqa: PLR0915
     audio_transcription_file_duration: float = 0.0,  # for audio transcription calls - the file time in seconds
     ### SERVICE TIER ###
     service_tier: Optional[str] = None,  # for OpenAI service tier pricing
+    custom_pricing: Optional[bool] = None,  # whether model uses custom pricing from router/config
     response: Optional[Any] = None,
     ### REQUEST MODEL ###
     request_model: Optional[str] = None,  # original request model for router detection
@@ -475,6 +476,13 @@ def cost_per_token(  # noqa: PLR0915
                 else None
             ),
         )
+    elif custom_pricing is True and model in model_cost_ref:
+        return generic_cost_per_token(
+            model=model,
+            usage=usage_block,
+            custom_llm_provider=custom_llm_provider or "",
+            service_tier=service_tier,
+        )
     elif custom_llm_provider == "vertex_ai":
         cost_router = google_cost_router(
             model=model_without_prefix,
@@ -545,8 +553,8 @@ def cost_per_token(  # noqa: PLR0915
         )
 
         if (
-            (model_info.get("input_cost_per_token") or 0.0) > 0
-            or (model_info.get("output_cost_per_token") or 0.0) > 0
+            model_info.get("input_cost_per_token") is not None
+            or model_info.get("output_cost_per_token") is not None
         ):
             return generic_cost_per_token(
                 model=model,
@@ -1509,6 +1517,7 @@ def completion_cost(  # noqa: PLR0915
                     audio_transcription_file_duration=audio_transcription_file_duration,
                     rerank_billed_units=rerank_billed_units,
                     service_tier=service_tier,
+                    custom_pricing=custom_pricing,
                     response=completion_response,
                     request_model=request_model_for_cost,
                 )

tests/test_litellm/test_cost_calculator.py

@@ -447,6 +447,77 @@ def test_per_request_custom_pricing_with_router():
     assert "gpt-3.5-turbo" in selected
 
 
+def test_zero_cost_custom_pricing_with_anthropic_provider():
+    """
+    When a user sets input_cost_per_token=0.0 and output_cost_per_token=0.0
+    with custom_llm_provider="anthropic" via the router, the zero-cost
+    override must be respected. The provider-specific dispatch should
+    be bypassed so that real Anthropic pricing is NOT used.
+
+    Regression test for https://github.com/BerriAI/litellm/issues/25204
+    """
+    from litellm import Router
+
+    router = Router(
+        model_list=[
+            {
+                "model_name": "free-claude",
+                "litellm_params": {
+                    "model": "anthropic/claude-sonnet-4-20250514",
+                    "api_key": "test_api_key",
+                },
+                "model_info": {
+                    "id": "zero-cost-model-id",
+                    "input_cost_per_token": 0.0,
+                    "output_cost_per_token": 0.0,
+                },
+            },
+        ]
+    )
+
+    result = router.completion(
+        model="free-claude",
+        messages=[{"role": "user", "content": "Hello!"}],
+        mock_response="Hi there!",
+    )
+
+    assert result._hidden_params["response_cost"] == 0.0
+
+
+def test_cost_per_token_custom_pricing_bypasses_provider_dispatch():
+    """
+    cost_per_token with custom_pricing=True should use generic_cost_per_token
+    instead of the provider-specific dispatch, ensuring custom pricing entries
+    (including zero costs) are respected.
+
+    Regression test for https://github.com/BerriAI/litellm/issues/25204
+    """
+    from litellm.cost_calculator import cost_per_token
+
+    custom_model_id = "custom-zero-cost-anthropic"
+    litellm.register_model(
+        model_cost={
+            custom_model_id: {
+                "input_cost_per_token": 0.0,
+                "output_cost_per_token": 0.0,
+                "litellm_provider": "anthropic",
+                "max_tokens": 8192,
+            }
+        }
+    )
+
+    prompt_cost, completion_cost_val = cost_per_token(
+        model=custom_model_id,
+        prompt_tokens=1000,
+        completion_tokens=500,
+        custom_llm_provider="anthropic",
+        custom_pricing=True,
+    )
+
+    assert prompt_cost == 0.0
+    assert completion_cost_val == 0.0
+
+
 def test_azure_realtime_cost_calculator():
     os.environ["LITELLM_LOCAL_MODEL_COST_MAP"] = "True"
     litellm.model_cost = litellm.get_model_cost_map(url="")