fix(openrouter): preserve provider-reported embedding cost

Follow the existing OpenRouter chat implementation by extracting usage.cost
from the raw response after conversion and recording it in provider-cost
metadata. Use set_response_cost_in_hidden_params so the cost calculator
can prioritize the reported cost over the local price map

This fixes zero spend for unregistered embedding models such as
openrouter/qwen/qwen3-embedding-8b. Preserve zero as a valid cost and retain
price-map fallback when cost is missing or invalid. Embeddings already
include cost without the usage.include request option required by chat

Supersedes #36891 with a focused change covering usage.cost only, without
new response schemas, cost_details handling, or cost-calculator changes

Signed-off-by: Viktor Oreshkin <solteke.st@gmail.com>

litellm/llms/openrouter/embedding/transformation.py

@@ -7,10 +7,12 @@ OpenRouter is OpenAI-compatible and supports embeddings via the /v1/embeddings e
 Docs: https://openrouter.ai/docs
 """
 
+import math
 from typing import TYPE_CHECKING, Any, Final
 
 import httpx
 
+from litellm.litellm_core_utils.core_helpers import set_response_cost_in_hidden_params
 from litellm.litellm_core_utils.url_utils import join_api_base
 from litellm.llms.base_llm.embedding.transformation import BaseEmbeddingConfig
 from litellm.types.llms.openai import AllEmbeddingInputValues
@@ -131,11 +133,20 @@ class OpenrouterEmbeddingConfig(BaseEmbeddingConfig):
         # OpenRouter returns standard OpenAI-compatible embedding response
         response_json: Final = raw_response.json()
 
-        return convert_to_model_response_object(
+        convert_to_model_response_object(
             response_object=response_json,
             model_response_object=model_response,
             response_type="embedding",
         )
+        response_cost: Final[object] = (response_json.get("usage") or {}).get("cost")
+        if (
+            isinstance(response_cost, int | float)
+            and not isinstance(response_cost, bool)
+            and response_cost >= 0
+            and math.isfinite(response_cost)
+        ):
+            set_response_cost_in_hidden_params(model_response, response_cost)
+        return model_response
 
     def get_supported_openai_params(self, model: str) -> list:
         """

tests/unit/llms/openrouter/test_openrouter_embedding_transformation.py

@@ -4,16 +4,105 @@ Unit tests for OpenRouter embedding transformation logic.
 
 import json
 from typing import Final
+from unittest.mock import Mock, patch
 
 import httpx
 import pytest
 import respx
 
 import litellm
+from litellm.cost_calculator import response_cost_calculator
 from litellm.llms.custom_httpx.http_handler import HTTPHandler
 from litellm.llms.openrouter.embedding.transformation import (
     OpenrouterEmbeddingConfig,
 )
+from litellm.types.utils import EmbeddingResponse
+
+
+def _embedding_response(usage: dict[str, object], model: str = "qwen/qwen3-embedding-8b") -> EmbeddingResponse:
+    raw_response: Final = httpx.Response(
+        200,
+        text=json.dumps(
+            {
+                "object": "list",
+                "data": [{"object": "embedding", "index": 0, "embedding": [0.1, 0.2]}],
+                "model": model,
+                "usage": usage,
+            }
+        ),
+    )
+    return OpenrouterEmbeddingConfig().transform_embedding_response(
+        model=f"openrouter/{model}",
+        raw_response=raw_response,
+        model_response=EmbeddingResponse(),
+        logging_obj=Mock(),
+        api_key="test-api-key",
+        request_data={},
+        optional_params={},
+        litellm_params={},
+    )
+
+
+@pytest.mark.parametrize("cost", [2e-8, 0])
+def test_openrouter_embedding_preserves_cost_without_price_map(cost: float) -> None:
+    with patch.dict(litellm.model_cost, {}, clear=True):
+        response: Final = _embedding_response({"prompt_tokens": 2, "total_tokens": 2, "cost": cost})
+        assert response._hidden_params["additional_headers"]["llm_provider-x-litellm-response-cost"] == cost
+        assert (
+            response_cost_calculator(
+                response_object=response,
+                model="openrouter/qwen/qwen3-embedding-8b",
+                custom_llm_provider="openrouter",
+                call_type="embedding",
+                optional_params={},
+            )
+            == cost
+        )
+    assert response.data[0]["embedding"] == [0.1, 0.2]
+    assert response.usage.prompt_tokens == response.usage.total_tokens == 2
+
+
+@pytest.mark.parametrize(
+    "cost_fields",
+    [
+        pytest.param({}, id="missing"),
+        pytest.param({"cost": None}, id="null"),
+        pytest.param({"cost": "invalid"}, id="string"),
+        pytest.param({"cost": True}, id="boolean"),
+        pytest.param({"cost": -1}, id="negative"),
+        pytest.param({"cost": {}}, id="object"),
+        pytest.param({"cost": float("nan")}, id="nan"),
+        pytest.param({"cost": float("inf")}, id="infinity"),
+    ],
+)
+@patch.dict(
+    litellm.model_cost,
+    {
+        "openrouter/openai/text-embedding-3-small": {
+            "input_cost_per_token": 0.001,
+            "output_cost_per_token": 0,
+            "litellm_provider": "openrouter",
+            "mode": "embedding",
+        }
+    },
+)
+def test_openrouter_embedding_missing_or_invalid_cost_uses_price_map(cost_fields: dict[str, object]) -> None:
+    response: Final = _embedding_response(
+        {"prompt_tokens": 2, "total_tokens": 2, **cost_fields}, model="openai/text-embedding-3-small"
+    )
+    assert "llm_provider-x-litellm-response-cost" not in response._hidden_params.get("additional_headers", {})
+    assert response.data[0]["embedding"] == [0.1, 0.2]
+    assert response.usage.prompt_tokens == response.usage.total_tokens == 2
+    assert (
+        response_cost_calculator(
+            response_object=response,
+            model="openai/text-embedding-3-small",
+            custom_llm_provider="openrouter",
+            call_type="embedding",
+            optional_params={},
+        )
+        == 2 * 0.001
+    )
 
 
 def test_openrouter_embedding_supported_params():