fix(proxy): forward resolved provider and deployment pricing in /cost/estimate

estimate_cost resolved on-prem aliases (e.g. nvidia/zai-org/glm-5.2) to their
underlying model and custom_llm_provider via the router, then called
completion_cost without either, so provider inference ran on the bare model and
raised "LLM Provider NOT provided"; deployment-configured per-token pricing was
dropped too, so priced on-prem deployments estimated 0. The resolver now returns
a frozen ResolvedCostModel(model, provider, custom_cost_per_token) and
estimate_cost forwards both into completion_cost and surfaces the configured
per-token pricing in the response, deriving that pricing as single Final values.

Resolves LIT-5210

litellm/proxy/management_endpoints/cost_tracking_settings.py

@@ -10,6 +10,8 @@ PATCH /config/cost_margin_config - Update cost margin configuration
 POST /cost/estimate - Estimate cost for a given model and token counts
 """
 
+from collections.abc import Mapping
+from dataclasses import dataclass
 from typing import Final
 
 from fastapi import APIRouter, Depends, HTTPException
@@ -24,29 +26,57 @@ from litellm.proxy._types import (
     UserAPIKeyAuth,
 )
 from litellm.proxy.auth.user_api_key_auth import user_api_key_auth
-from litellm.types.utils import LlmProvidersSet
+from litellm.types.utils import CostPerToken, LlmProvidersSet, ModelInfo
 
 router: Final = APIRouter()
 
 
-def _resolve_model_for_cost_lookup(model: str) -> tuple[str, str | None]:
+@dataclass(frozen=True, slots=True)
+class ResolvedCostModel:
+    model: str
+    provider: str | None
+    custom_cost_per_token: CostPerToken | None
+
+
+def _extract_custom_pricing(litellm_params: Mapping[str, object]) -> CostPerToken | None:
+    """
+    Pull per-token pricing configured on a deployment so on-prem / self-hosted
+    models (absent from the public cost map) still estimate a real cost.
+    """
+    input_cost: Final = litellm_params.get("input_cost_per_token")
+    output_cost: Final = litellm_params.get("output_cost_per_token")
+
+    input_price: Final = float(input_cost) if isinstance(input_cost, (int, float)) else None
+    output_price: Final = float(output_cost) if isinstance(output_cost, (int, float)) else None
+
+    if input_price is None and output_price is None:
+        return None
+
+    return CostPerToken(
+        input_cost_per_token=input_price or 0.0,
+        output_cost_per_token=output_price or 0.0,
+    )
+
+
+def _lookup_model_info(model: str) -> ModelInfo | None:
+    try:
+        return litellm.get_model_info(model=model)
+    except Exception:
+        return None
+
+
+def _resolve_model_for_cost_lookup(model: str) -> ResolvedCostModel:
     """
     Resolve a model name (which may be a router alias/model_group) to the
-    underlying litellm model name for cost lookup.
+    underlying litellm model name, provider, and any deployment-configured
+    pricing used for cost lookup.
 
     Args:
         model: The model name from the request (could be a router alias like 'e-model-router'
                or an actual model name like 'azure_ai/gpt-4')
-
-    Returns:
-        Tuple of (resolved_model_name, custom_llm_provider)
-        - resolved_model_name: The actual model name to use for cost lookup
-        - custom_llm_provider: The provider if resolved from router, None otherwise
     """
     from litellm.proxy.proxy_server import llm_router
 
-    custom_llm_provider: str | None = None
-
     # Try to resolve from router if available
     if llm_router is not None:
         try:
@@ -57,31 +87,25 @@ def _resolve_model_for_cost_lookup(model: str) -> tuple[str, str | None]:
                 first_deployment: Final = deployments[0]
                 litellm_params: Final = first_deployment.get("litellm_params", {})
                 model_info: Final = first_deployment.get("model_info", {})
+                custom_llm_provider: Final = litellm_params.get("custom_llm_provider")
+                provider: Final = str(custom_llm_provider) if custom_llm_provider is not None else None
+                custom_cost_per_token: Final = _extract_custom_pricing(litellm_params)
 
                 # Check base_model first (needed for Azure custom deployment names)
                 base_model: Final = model_info.get("base_model") or litellm_params.get("base_model")
                 if base_model:
                     verbose_proxy_logger.debug("Resolved model '%s' to base_model '%s' from router", model, base_model)
-                    custom_llm_provider = litellm_params.get("custom_llm_provider")
-                    return (
-                        str(base_model),
-                        (str(custom_llm_provider) if custom_llm_provider is not None else None),
-                    )
+                    return ResolvedCostModel(str(base_model), provider, custom_cost_per_token)
 
                 resolved_model: Final = litellm_params.get("model")
-
                 if resolved_model:
                     verbose_proxy_logger.debug("Resolved model '%s' to '%s' from router", model, resolved_model)
-                    custom_llm_provider = litellm_params.get("custom_llm_provider")
-                    return (
-                        str(resolved_model),
-                        (str(custom_llm_provider) if custom_llm_provider is not None else None),
-                    )
+                    return ResolvedCostModel(str(resolved_model), provider, custom_cost_per_token)
         except Exception as e:
             verbose_proxy_logger.debug("Could not resolve model '%s' from router: %s", model, e)
 
     # Return original model if not resolved
-    return model, custom_llm_provider
+    return ResolvedCostModel(model, None, None)
 
 
 def _calculate_period_costs(num_requests, cost_per_request, input_cost, output_cost, margin_cost):
@@ -450,7 +474,9 @@ async def estimate_cost(
     from litellm.types.utils import ModelResponse, Usage
 
     # Resolve model name (handles router aliases like 'e-model-router' -> 'azure_ai/gpt-4')
-    resolved_model, resolved_provider = _resolve_model_for_cost_lookup(request.model)
+    resolved: Final = _resolve_model_for_cost_lookup(request.model)
+    resolved_model: Final = resolved.model
+    resolved_provider: Final = resolved.provider
 
     verbose_proxy_logger.debug("Cost estimate: request.model='%s' resolved to '%s'", request.model, resolved_model)
 
@@ -480,6 +506,8 @@ async def estimate_cost(
         cost_per_request: Final = completion_cost(
             completion_response=mock_response,
             model=resolved_model,
+            custom_llm_provider=resolved_provider,
+            custom_cost_per_token=resolved.custom_cost_per_token,
             litellm_logging_obj=litellm_logging_obj,
         )
     except Exception as e:
@@ -497,20 +525,22 @@ async def estimate_cost(
     output_cost: Final = cost_breakdown.get("output_cost", 0.0) if cost_breakdown else 0.0
     margin_cost: Final = cost_breakdown.get("margin_total_amount", 0.0) if cost_breakdown else 0.0
 
-    # Get model info for per-token pricing display
-    try:
-        model_info: Final = litellm.get_model_info(model=resolved_model)
-        input_cost_per_token = model_info.get("input_cost_per_token")
-        output_cost_per_token = model_info.get("output_cost_per_token")
-        custom_llm_provider = model_info.get("litellm_provider")
-    except Exception:
-        input_cost_per_token = None
-        output_cost_per_token = None
-        custom_llm_provider = None
+    model_info: Final = _lookup_model_info(resolved_model)
+    mapped_input_price: Final = model_info.get("input_cost_per_token") if model_info is not None else None
+    mapped_output_price: Final = model_info.get("output_cost_per_token") if model_info is not None else None
+    mapped_provider: Final = model_info.get("litellm_provider") if model_info is not None else None
 
-    # Use provider from router resolution if not found in model_info
-    if custom_llm_provider is None and resolved_provider is not None:
-        custom_llm_provider = resolved_provider
+    input_cost_per_token: Final = (
+        resolved.custom_cost_per_token["input_cost_per_token"]
+        if resolved.custom_cost_per_token is not None
+        else mapped_input_price
+    )
+    output_cost_per_token: Final = (
+        resolved.custom_cost_per_token["output_cost_per_token"]
+        if resolved.custom_cost_per_token is not None
+        else mapped_output_price
+    )
+    custom_llm_provider: Final = mapped_provider if mapped_provider is not None else resolved_provider
 
     # Calculate daily and monthly costs
     (

tests/test_litellm/proxy/management_endpoints/test_cost_tracking_settings.py

@@ -322,9 +322,9 @@ class TestResolveModelForCostLookup:
             "litellm.proxy.proxy_server.llm_router",
             mock_router,
         ):
-            resolved_model, provider = _resolve_model_for_cost_lookup("gpt-5.3-codex")
+            resolved = _resolve_model_for_cost_lookup("gpt-5.3-codex")
 
-        assert resolved_model == "azure/gpt-4o"
+        assert resolved.model == "azure/gpt-4o"
         mock_router.get_model_list.assert_called_once_with(model_name="gpt-5.3-codex")
 
     def test_falls_back_to_litellm_params_model_when_no_base_model(self):
@@ -352,9 +352,9 @@ class TestResolveModelForCostLookup:
             "litellm.proxy.proxy_server.llm_router",
             mock_router,
         ):
-            resolved_model, provider = _resolve_model_for_cost_lookup("gpt-4")
+            resolved = _resolve_model_for_cost_lookup("gpt-4")
 
-        assert resolved_model == "openai/gpt-4"
+        assert resolved.model == "openai/gpt-4"
 
     def test_resolves_base_model_from_litellm_params(self):
         """
@@ -383,9 +383,9 @@ class TestResolveModelForCostLookup:
             "litellm.proxy.proxy_server.llm_router",
             mock_router,
         ):
-            resolved_model, provider = _resolve_model_for_cost_lookup("my-azure-model")
+            resolved = _resolve_model_for_cost_lookup("my-azure-model")
 
-        assert resolved_model == "azure/gpt-4o-mini"
+        assert resolved.model == "azure/gpt-4o-mini"
 
     def test_returns_original_model_when_no_router(self):
         """
@@ -399,12 +399,10 @@ class TestResolveModelForCostLookup:
             "litellm.proxy.proxy_server.llm_router",
             None,
         ):
-            resolved_model, provider = _resolve_model_for_cost_lookup(
-                "azure/openai/gpt-5.3-codex"
-            )
+            resolved = _resolve_model_for_cost_lookup("azure/openai/gpt-5.3-codex")
 
-        assert resolved_model == "azure/openai/gpt-5.3-codex"
-        assert provider is None
+        assert resolved.model == "azure/openai/gpt-5.3-codex"
+        assert resolved.provider is None
 
     def test_returns_custom_llm_provider_on_base_model_path(self):
         """base_model path: the custom_llm_provider from litellm_params is
@@ -427,10 +425,10 @@ class TestResolveModelForCostLookup:
         ]
 
         with patch("litellm.proxy.proxy_server.llm_router", mock_router):
-            resolved_model, provider = _resolve_model_for_cost_lookup("my-azure-model")
+            resolved = _resolve_model_for_cost_lookup("my-azure-model")
 
-        assert resolved_model == "azure/gpt-4o"
-        assert provider == "azure"
+        assert resolved.model == "azure/gpt-4o"
+        assert resolved.provider == "azure"
 
     def test_returns_custom_llm_provider_on_resolved_model_path(self):
         """resolved-model path (no base_model): the custom_llm_provider from
@@ -452,10 +450,10 @@ class TestResolveModelForCostLookup:
         ]
 
         with patch("litellm.proxy.proxy_server.llm_router", mock_router):
-            resolved_model, provider = _resolve_model_for_cost_lookup("gpt-4")
+            resolved = _resolve_model_for_cost_lookup("gpt-4")
 
-        assert resolved_model == "openai/gpt-4"
-        assert provider == "openai"
+        assert resolved.model == "openai/gpt-4"
+        assert resolved.provider == "openai"
 
     def test_resolves_base_model_when_deployment_has_no_litellm_params(self):
         """A deployment can omit litellm_params entirely; base_model from
@@ -474,10 +472,10 @@ class TestResolveModelForCostLookup:
         ]
 
         with patch("litellm.proxy.proxy_server.llm_router", mock_router):
-            resolved_model, provider = _resolve_model_for_cost_lookup("my-azure-model")
+            resolved = _resolve_model_for_cost_lookup("my-azure-model")
 
-        assert resolved_model == "azure/gpt-4o"
-        assert provider is None
+        assert resolved.model == "azure/gpt-4o"
+        assert resolved.provider is None
 
     def test_resolves_model_when_deployment_has_no_model_info(self):
         """A deployment can omit model_info entirely; litellm_params.model must
@@ -496,7 +494,109 @@ class TestResolveModelForCostLookup:
         ]
 
         with patch("litellm.proxy.proxy_server.llm_router", mock_router):
-            resolved_model, provider = _resolve_model_for_cost_lookup("gpt-4")
+            resolved = _resolve_model_for_cost_lookup("gpt-4")
 
-        assert resolved_model == "openai/gpt-4"
-        assert provider is None
+        assert resolved.model == "openai/gpt-4"
+        assert resolved.provider is None
+
+
+class TestEstimateCostOnPremProvider:
+    """Regression tests for LIT-5210: /cost/estimate on on-prem deployment aliases."""
+
+    @pytest.mark.asyncio
+    async def test_estimate_cost_onprem_model_without_pricing(self):
+        """
+        On-prem deployments (custom_llm_provider set, model absent from the cost map)
+        must not 500 with "LLM Provider NOT provided". The resolved provider has to be
+        forwarded to completion_cost so provider inference doesn't run on the bare model.
+
+        completion_cost is intentionally NOT mocked.
+        """
+        from litellm.proxy._types import CostEstimateRequest
+        from litellm.proxy.management_endpoints.cost_tracking_settings import (
+            estimate_cost,
+        )
+
+        request = CostEstimateRequest(
+            model="nvidia/zai-org/glm-5.2",
+            input_tokens=1000,
+            output_tokens=500,
+        )
+
+        mock_router = MagicMock()
+        mock_router.get_model_list.return_value = [
+            {
+                "model_name": "nvidia/zai-org/glm-5.2",
+                "litellm_params": {
+                    "model": "zai-org/GLM-5.2",
+                    "custom_llm_provider": "openai",
+                },
+                "model_info": {},
+            }
+        ]
+
+        saved_model_cost = dict(litellm.model_cost)
+        litellm.register_model(
+            {
+                "openai/zai-org/GLM-5.2": {
+                    "input_cost_per_token": 0.0,
+                    "output_cost_per_token": 0.0,
+                    "litellm_provider": "openai",
+                    "mode": "chat",
+                }
+            }
+        )
+        try:
+            with patch("litellm.proxy.proxy_server.llm_router", mock_router):
+                response = await estimate_cost(request=request, user_api_key_dict=MagicMock())
+        finally:
+            litellm.model_cost = saved_model_cost
+
+        assert response.model == "nvidia/zai-org/glm-5.2"
+        assert response.provider == "openai"
+        assert response.cost_per_request == 0.0
+
+    @pytest.mark.asyncio
+    async def test_estimate_cost_onprem_model_with_configured_pricing(self):
+        """
+        On-prem deployments with input/output_cost_per_token configured must estimate a
+        real cost using that pricing, not fall back to 0.0.
+
+        completion_cost is intentionally NOT mocked.
+        """
+        from litellm.proxy._types import CostEstimateRequest
+        from litellm.proxy.management_endpoints.cost_tracking_settings import (
+            estimate_cost,
+        )
+
+        request = CostEstimateRequest(
+            model="nvidia/zai-org/glm-5.2",
+            input_tokens=1000,
+            output_tokens=500,
+            num_requests_per_day=100,
+        )
+
+        mock_router = MagicMock()
+        mock_router.get_model_list.return_value = [
+            {
+                "model_name": "nvidia/zai-org/glm-5.2",
+                "litellm_params": {
+                    "model": "zai-org/GLM-5.2",
+                    "custom_llm_provider": "openai",
+                    "input_cost_per_token": 0.000001,
+                    "output_cost_per_token": 0.000002,
+                },
+                "model_info": {},
+            }
+        ]
+
+        with patch("litellm.proxy.proxy_server.llm_router", mock_router):
+            response = await estimate_cost(request=request, user_api_key_dict=MagicMock())
+
+        assert response.provider == "openai"
+        assert response.cost_per_request == pytest.approx(0.002)
+        assert response.input_cost_per_request == pytest.approx(0.001)
+        assert response.output_cost_per_request == pytest.approx(0.001)
+        assert response.daily_cost == pytest.approx(0.2)
+        assert response.input_cost_per_token == pytest.approx(0.000001)
+        assert response.output_cost_per_token == pytest.approx(0.000002)