fix(search): forward search-tool params through the router, complete Parallel AI v1 param mapping

SearchAPIRouter dropped every parameter configured on a search tool, forwarding
only per-request kwargs. Any tool-level setting (mode, max_results, ...) was
silently lost on the way to the adapter, for every search provider.

Also completes the Parallel AI v1 search surface: after_date, fetch_policy,
location and include_domains now nest under advanced_settings instead of being
sent as unknown top-level fields, responses preserve search_id / session_id /
warnings / raw excerpts, and search cost is derived from the request mode and
the provider's reported usage rather than a single flat rate.

litellm/llms/parallel_ai/search/cost_calculator.py

@@ -0,0 +1,82 @@
+from collections.abc import Mapping, Sequence
+from typing import Final
+
+from pydantic import TypeAdapter, ValidationError
+
+from litellm.utils import get_model_info
+
+PARALLEL_AI_DEFAULT_RESULTS: Final = 10
+PARALLEL_AI_ADDITIONAL_RESULT_COST: Final = 0.001
+PARALLEL_AI_USAGE_PARAM: Final = "_parallel_ai_usage"
+PARALLEL_AI_STANDARD_SEARCH_MODEL: Final = "parallel_ai/search"
+PARALLEL_AI_TURBO_SEARCH_MODEL: Final = "parallel_ai/search-turbo"
+ADVANCED_SETTINGS_ADAPTER: Final[TypeAdapter[Mapping[str, object]]] = TypeAdapter(Mapping[str, object])
+
+
+def _non_negative_int(value: object) -> int | None:
+    if isinstance(value, bool) or not isinstance(value, int) or value < 0:
+        return None
+    return value
+
+
+def _usage_count(usage: Sequence[Mapping[str, object]], sku: str) -> int | None:
+    counts: Final = tuple(
+        count
+        for item in usage
+        if item.get("name") == sku
+        if (count := _non_negative_int(item.get("count"))) is not None
+    )
+    return sum(counts) if counts else None
+
+
+def _effective_mode(optional_params: Mapping[str, object]) -> str:
+    mode: Final = optional_params.get("mode")
+    if isinstance(mode, str):
+        return mode
+
+    processor: Final = optional_params.get("processor")
+    if processor == "pro":
+        return "advanced"
+    return "basic"
+
+
+def _effective_max_results(optional_params: Mapping[str, object]) -> int:
+    try:
+        advanced_settings: Final = ADVANCED_SETTINGS_ADAPTER.validate_python(optional_params.get("advanced_settings"))
+        advanced_max_results: Final = _non_negative_int(advanced_settings.get("max_results"))
+        if advanced_max_results is not None:
+            return advanced_max_results
+    except ValidationError:
+        pass
+
+    max_results: Final = _non_negative_int(optional_params.get("max_results"))
+    return max_results if max_results is not None else PARALLEL_AI_DEFAULT_RESULTS
+
+
+def _request_cost(mode: str) -> float:
+    pricing_model: Final = PARALLEL_AI_TURBO_SEARCH_MODEL if mode == "turbo" else PARALLEL_AI_STANDARD_SEARCH_MODEL
+    model_info: Final = get_model_info(model=pricing_model, custom_llm_provider="parallel_ai")
+    return float(model_info.get("input_cost_per_query") or 0.0)
+
+
+def _additional_results(
+    optional_params: Mapping[str, object],
+    usage: Sequence[Mapping[str, object]] | None,
+) -> int:
+    usage_count: Final = _usage_count(usage, "sku_search_additional_results") if usage is not None else None
+    if usage_count is not None:
+        return usage_count
+    if usage is not None:
+        return 0
+    return max(_effective_max_results(optional_params) - PARALLEL_AI_DEFAULT_RESULTS, 0)
+
+
+def parallel_ai_search_cost(
+    optional_params: Mapping[str, object],
+    usage: Sequence[Mapping[str, object]] | None,
+) -> float:
+    request_cost: Final = _request_cost(_effective_mode(optional_params))
+    request_count_from_usage: Final = _usage_count(usage, "sku_search") if usage is not None else None
+    request_count: Final = request_count_from_usage if request_count_from_usage is not None else 1
+    additional_results: Final = _additional_results(optional_params, usage)
+    return request_count * request_cost + additional_results * PARALLEL_AI_ADDITIONAL_RESULT_COST

litellm/llms/parallel_ai/search/transformation.py

@@ -4,9 +4,13 @@ Calls Parallel AI's /v1/search endpoint to search the web.
 Parallel AI API Reference: https://docs.parallel.ai/api-reference/search/search
 """
 
+from collections.abc import Mapping, Sequence
+from types import MappingProxyType
 from typing import Final, TypedDict
 
 import httpx
+from pydantic import BaseModel, ConfigDict
+from typing_extensions import ReadOnly
 
 from litellm.litellm_core_utils.litellm_logging import Logging as LiteLLMLoggingObj
 from litellm.llms.base_llm.search.transformation import (
@@ -14,9 +18,29 @@ from litellm.llms.base_llm.search.transformation import (
     SearchResponse,
     SearchResult,
 )
+from litellm.llms.parallel_ai.search.cost_calculator import PARALLEL_AI_USAGE_PARAM
 from litellm.secret_managers.main import get_secret_str
 
 
+class _ParallelAIV1SearchResult(BaseModel):
+    model_config = ConfigDict(extra="ignore")
+
+    url: str = ""
+    title: str | None = None
+    publish_date: str | None = None
+    excerpts: Sequence[str] = ()
+
+
+class _ParallelAIV1SearchResponse(BaseModel):
+    model_config = ConfigDict(extra="ignore")
+
+    search_id: str | None = None
+    session_id: str | None = None
+    results: Sequence[_ParallelAIV1SearchResult] = ()
+    usage: Sequence[Mapping[str, object]] | None = None
+    warnings: Sequence[Mapping[str, object]] | None = None
+
+
 class _ParallelAISourcePolicy(TypedDict, total=False):
     include_domains: list[str]
     exclude_domains: list[str]
@@ -27,10 +51,16 @@ class _ParallelAIExcerptSettings(TypedDict, total=False):
     max_chars_per_result: int
 
 
+class _ParallelAIFetchPolicy(TypedDict, total=False):
+    max_age_seconds: ReadOnly[int]
+    timeout_seconds: ReadOnly[float]
+    disable_cache_fallback: ReadOnly[bool]
+
+
 class _ParallelAIAdvancedSettings(TypedDict, total=False):
     source_policy: _ParallelAISourcePolicy
     excerpt_settings: _ParallelAIExcerptSettings
-    fetch_policy: dict
+    fetch_policy: _ParallelAIFetchPolicy
     location: str
     max_results: int
 
@@ -50,7 +80,7 @@ class ParallelAISearchRequest(TypedDict, total=False):
     advanced_settings: _ParallelAIAdvancedSettings
 
 
-LEGACY_PROCESSOR_TO_MODE: Final = {"base": "basic", "pro": "advanced"}
+LEGACY_PROCESSOR_TO_MODE: Final = MappingProxyType({"base": "basic", "pro": "advanced"})
 
 
 class ParallelAISearchConfig(BaseSearchConfig):
@@ -67,16 +97,16 @@ class ParallelAISearchConfig(BaseSearchConfig):
         api_base: str | None = None,
         **kwargs,
     ) -> dict:
-        api_key = self.resolve_server_api_key(
+        resolved_api_key: Final = self.resolve_server_api_key(
             caller_api_key=api_key,
             caller_api_base=api_base,
             key_env_vars=("PARALLEL_AI_API_KEY", "PARALLEL_API_KEY"),
             base_env_var="PARALLEL_AI_API_BASE",
             default_api_base=self.PARALLEL_AI_API_BASE,
         )
-        if not api_key:
+        if not resolved_api_key:
             raise ValueError("PARALLEL_API_KEY is not set. Set `PARALLEL_API_KEY` environment variable.")
-        headers["x-api-key"] = api_key
+        headers["x-api-key"] = resolved_api_key
         headers["Content-Type"] = "application/json"
         return headers
 
@@ -87,13 +117,12 @@ class ParallelAISearchConfig(BaseSearchConfig):
         data: dict | list[dict] | None = None,
         **kwargs,
     ) -> str:
-        api_base = api_base or get_secret_str("PARALLEL_AI_API_BASE") or self.PARALLEL_AI_API_BASE
+        resolved_api_base: Final = api_base or get_secret_str("PARALLEL_AI_API_BASE") or self.PARALLEL_AI_API_BASE
 
-        api_base = api_base.rstrip("/")
-        if not api_base.endswith("/v1/search"):
-            api_base = f"{api_base.removesuffix('/v1')}/v1/search"
-
-        return api_base
+        trimmed: Final = resolved_api_base.rstrip("/")
+        if trimmed.endswith("/v1/search"):
+            return trimmed
+        return f"{trimmed.removesuffix('/v1')}/v1/search"
 
     def transform_search_request(
         self,
@@ -112,11 +141,14 @@ class ParallelAISearchConfig(BaseSearchConfig):
                 - mode: Search mode ('turbo', 'basic', 'advanced'); defaults to 'basic'
                 - processor: Legacy v1beta param; 'base' maps to mode 'basic', 'pro' to 'advanced'
                 - max_results: Maximum number of search results -> `advanced_settings.max_results`
-                - search_domain_filter: Domains to include -> `advanced_settings.source_policy.include_domains`
+                - search_domain_filter / include_domains: Domains to include -> `advanced_settings.source_policy.include_domains`
                 - exclude_domains: Domains to exclude -> `advanced_settings.source_policy.exclude_domains`
-                - country: ISO 3166-1 alpha-2 code -> `advanced_settings.location`
+                - after_date: RFC 3339 date (YYYY-MM-DD) -> `advanced_settings.source_policy.after_date`
+                - country / location: ISO 3166-1 alpha-2 code -> `advanced_settings.location`
                 - max_chars_per_result: -> `advanced_settings.excerpt_settings.max_chars_per_result`
-                - Any other params are passed through to the request body as-is
+                - fetch_policy: Cache vs live-fetch policy -> `advanced_settings.fetch_policy`
+                - Any other params (objective, max_chars_total, session_id, client_model, ...)
+                  are passed through to the request body as-is
 
         Returns:
             Dict with request data following the v1 search request spec
@@ -137,7 +169,7 @@ class ParallelAISearchConfig(BaseSearchConfig):
             mode = LEGACY_PROCESSOR_TO_MODE.get(processor, processor)
         # the v1 API defaults to 'advanced' when mode is omitted; default to 'basic'
         # instead to keep v1beta's default tier (processor 'base') and litellm's
-        # $0.004/query cost map entry for `parallel_ai/search` accurate
+        # cost map entry for `parallel_ai/search` accurate
         request_data["mode"] = mode or "basic"
 
         advanced_settings: Final[_ParallelAIAdvancedSettings] = {}
@@ -148,17 +180,29 @@ class ParallelAISearchConfig(BaseSearchConfig):
         if "country" in params:
             advanced_settings["location"] = params.pop("country")
 
+        if "location" in params:
+            advanced_settings["location"] = params.pop("location")
+
         if "max_chars_per_result" in params:
             advanced_settings["excerpt_settings"] = {"max_chars_per_result": params.pop("max_chars_per_result")}
 
+        if "fetch_policy" in params:
+            advanced_settings["fetch_policy"] = params.pop("fetch_policy")
+
         source_policy: Final[_ParallelAISourcePolicy] = {}
 
         if "search_domain_filter" in params:
             source_policy["include_domains"] = params.pop("search_domain_filter")
 
+        if "include_domains" in params:
+            source_policy["include_domains"] = params.pop("include_domains")
+
         if "exclude_domains" in params:
             source_policy["exclude_domains"] = params.pop("exclude_domains")
 
+        if "after_date" in params:
+            source_policy["after_date"] = params.pop("after_date")
+
         if source_policy:
             advanced_settings["source_policy"] = source_policy
 
@@ -170,9 +214,7 @@ class ParallelAISearchConfig(BaseSearchConfig):
         # unified-spec param with no v1 equivalent
         params.pop("max_tokens_per_page", None)
 
-        result_data: Final[dict] = dict(request_data)
-        result_data.update(params)
-        return result_data
+        return {**request_data, **params}
 
     def transform_search_response(
         self,
@@ -186,26 +228,48 @@ class ParallelAISearchConfig(BaseSearchConfig):
         Parallel AI -> LiteLLM mappings:
         - results[].title -> SearchResult.title
         - results[].url -> SearchResult.url
-        - results[].excerpts (array) -> SearchResult.snippet (joined string)
+        - results[].excerpts (array) -> SearchResult.snippet (joined string); the raw
+          array is preserved as an extra `excerpts` field on each result
         - results[].publish_date -> SearchResult.date
+        - search_id / session_id / warnings are preserved as extra fields on the
+          response; usage is preserved as `parallel_usage` (the `usage` name is
+          reserved for LiteLLM's token-usage object)
         """
-        response_json: Final = raw_response.json()
+        parsed: Final = _ParallelAIV1SearchResponse.model_validate(raw_response.json())
 
-        results: Final = []
-        for result in response_json.get("results", []):
-            excerpts = result.get("excerpts") or []
-            snippet = " ... ".join(excerpts) if excerpts else ""
+        if parsed.usage is not None:
+            logging_obj.optional_params = {
+                **logging_obj.optional_params,
+                PARALLEL_AI_USAGE_PARAM: parsed.usage,
+            }
 
-            search_result = SearchResult(
-                title=result.get("title") or "",
-                url=result.get("url") or "",
-                snippet=snippet,
-                date=result.get("publish_date"),
-                last_updated=None,
+        results: Final = tuple(
+            SearchResult.model_validate(
+                MappingProxyType(
+                    {
+                        "title": result.title or "",
+                        "url": result.url,
+                        "snippet": " ... ".join(result.excerpts) if result.excerpts else "",
+                        "date": result.publish_date,
+                        "last_updated": None,
+                        "excerpts": result.excerpts,
+                    }
+                )
             )
-            results.append(search_result)
-
-        return SearchResponse(
-            results=results,
-            object="search",
+            for result in parsed.results
         )
+
+        extra_fields: Final = MappingProxyType(
+            {
+                key: value
+                for key, value in (
+                    ("search_id", parsed.search_id),
+                    ("session_id", parsed.session_id),
+                    ("parallel_usage", parsed.usage),
+                    ("warnings", parsed.warnings),
+                )
+                if value is not None
+            }
+        )
+
+        return SearchResponse.model_validate(MappingProxyType({"results": results, "object": "search", **extra_fields}))

litellm/model_prices_and_context_window_backup.json

@@ -35283,12 +35283,17 @@
         "source": "https://cloud.google.com/vertex-ai/generative-ai/docs/learn/models#foundation_models"
     },
     "parallel_ai/search": {
-        "input_cost_per_query": 0.004,
+        "input_cost_per_query": 0.005,
         "litellm_provider": "parallel_ai",
         "mode": "search"
     },
     "parallel_ai/search-pro": {
-        "input_cost_per_query": 0.009,
+        "input_cost_per_query": 0.005,
+        "litellm_provider": "parallel_ai",
+        "mode": "search"
+    },
+    "parallel_ai/search-turbo": {
+        "input_cost_per_query": 0.001,
         "litellm_provider": "parallel_ai",
         "mode": "search"
     },

litellm/router_utils/search_api_router.py

@@ -9,6 +9,7 @@ import random
 import traceback
 from collections.abc import Callable
 from functools import partial
+from types import MappingProxyType
 from typing import Any, Final
 
 from litellm._logging import verbose_router_logger
@@ -214,6 +215,15 @@ class SearchAPIRouter:
             api_key, api_base = SearchAPIRouter._resolve_search_provider_credentials(
                 tool_litellm_params=litellm_params,
             )
+            protected_params: Final = frozenset(("search_provider", "api_key", "api_base"))
+            search_params: Final = MappingProxyType(
+                {
+                    key: value
+                    for params in (litellm_params, kwargs)
+                    for key, value in params.items()
+                    if key not in protected_params and value is not None
+                }
+            )
 
             verbose_router_logger.debug("Selected search tool with provider: %s", search_provider)
 
@@ -222,7 +232,7 @@ class SearchAPIRouter:
                 search_provider=search_provider,
                 api_key=api_key,
                 api_base=api_base,
-                **kwargs,
+                **search_params,
             )
 
             return response

litellm/search/cost_calculator.py

@@ -2,16 +2,37 @@
 Cost calculation for search providers.
 """
 
+from collections.abc import Mapping
+from types import MappingProxyType
 from typing import Final
 
+from pydantic import TypeAdapter, ValidationError
+
 from litellm.utils import get_model_info
 
+PROVIDER_USAGE_ADAPTER: Final[TypeAdapter[tuple[Mapping[str, object], ...]]] = TypeAdapter(
+    tuple[Mapping[str, object], ...]
+)
+EMPTY_OPTIONAL_PARAMS: Final[Mapping[str, object]] = MappingProxyType({})
+
+
+def _provider_usage(
+    optional_params: Mapping[str, object] | None,
+    usage_param: str,
+) -> tuple[Mapping[str, object], ...] | None:
+    params: Final = optional_params if optional_params is not None else EMPTY_OPTIONAL_PARAMS
+    raw_usage: Final[object] = params.get(usage_param)
+    try:
+        return PROVIDER_USAGE_ADAPTER.validate_python(raw_usage)
+    except ValidationError:
+        return None
+
 
 def search_provider_cost_per_query(
     model: str,
     custom_llm_provider: str | None = None,
     number_of_queries: int = 1,
-    optional_params: dict | None = None,
+    optional_params: Mapping[str, object] | None = None,
 ) -> tuple[float, float]:
     """
     Calculate cost for search-only providers.
@@ -28,6 +49,18 @@ def search_provider_cost_per_query(
     Returns:
         Tuple of (input_cost, output_cost) where output_cost is always 0.0
     """
+    if custom_llm_provider == "parallel_ai":
+        from litellm.llms.parallel_ai.search.cost_calculator import (
+            PARALLEL_AI_USAGE_PARAM,
+            parallel_ai_search_cost,
+        )
+
+        input_cost: Final = parallel_ai_search_cost(
+            optional_params=optional_params if optional_params is not None else EMPTY_OPTIONAL_PARAMS,
+            usage=_provider_usage(optional_params, PARALLEL_AI_USAGE_PARAM),
+        )
+        return (input_cost, 0.0)
+
     model_info: Final = get_model_info(model=model, custom_llm_provider=custom_llm_provider)
 
     # Check for tiered pricing (e.g., Exa AI based on max_results)

model_prices_and_context_window.json

@@ -35283,12 +35283,17 @@
         "source": "https://cloud.google.com/vertex-ai/generative-ai/docs/learn/models#foundation_models"
     },
     "parallel_ai/search": {
-        "input_cost_per_query": 0.004,
+        "input_cost_per_query": 0.005,
         "litellm_provider": "parallel_ai",
         "mode": "search"
     },
     "parallel_ai/search-pro": {
-        "input_cost_per_query": 0.009,
+        "input_cost_per_query": 0.005,
+        "litellm_provider": "parallel_ai",
+        "mode": "search"
+    },
+    "parallel_ai/search-turbo": {
+        "input_cost_per_query": 0.001,
         "litellm_provider": "parallel_ai",
         "mode": "search"
     },

tests/router_unit_tests/test_router_helper_utils.py

@@ -2311,6 +2311,7 @@ def search_tools():
                 "search_provider": "perplexity",
                 "api_key": "test-api-key",
                 "api_base": "https://api.perplexity.ai",
+                "mode": "turbo",
             },
         },
         {
@@ -2319,6 +2320,7 @@ def search_tools():
                 "search_provider": "perplexity",
                 "api_key": "test-api-key-2",
                 "api_base": "https://api.perplexity.ai",
+                "mode": "turbo",
             },
         },
     ]
@@ -2410,6 +2412,7 @@ async def test_asearch_with_fallbacks_helper(search_tools):
         assert "search_provider" in kwargs
         assert kwargs["search_provider"] == "perplexity"
         assert "api_key" in kwargs
+        assert kwargs["mode"] == "turbo"
         assert kwargs["query"] == "helper test query"
         return mock_response
 

tests/test_litellm/llms/parallel_ai/test_parallel_ai_search.py

@@ -33,13 +33,30 @@ MOCK_V1_RESPONSE = {
 }
 
 
-def _mock_response():
+def _mock_response(payload=None):
     mock_response = MagicMock()
     mock_response.status_code = 200
-    mock_response.json.return_value = MOCK_V1_RESPONSE
+    mock_response.json.return_value = payload if payload is not None else MOCK_V1_RESPONSE
     return mock_response
 
 
+
+@pytest.fixture
+def bundled_cost_map(monkeypatch):
+    """Price lookups against the bundled cost map.
+
+    litellm caches model-info lookups, so swapping ``model_cost`` only takes
+    effect once those caches are invalidated -- on the way in and back out.
+    """
+    from litellm.utils import _invalidate_model_cost_lowercase_map
+
+    monkeypatch.setattr(litellm, "model_cost", litellm.get_model_cost_map(url=""))
+    _invalidate_model_cost_lowercase_map()
+    yield
+    monkeypatch.undo()
+    _invalidate_model_cost_lowercase_map()
+
+
 class TestParallelAISearch:
     @pytest.fixture(autouse=True)
     def _set_api_key(self, monkeypatch):
@@ -341,3 +358,117 @@ class TestParallelAISearch:
                 query="AI developments",
                 search_provider="parallel_ai",
             )
+
+    @pytest.mark.asyncio
+    async def test_flat_source_and_fetch_params_nest_under_advanced_settings(self):
+        with patch(
+            "litellm.llms.custom_httpx.http_handler.AsyncHTTPHandler.post",
+            new_callable=AsyncMock,
+        ) as mock_post:
+            mock_post.return_value = _mock_response()
+
+            await litellm.asearch(
+                query="AI developments",
+                search_provider="parallel_ai",
+                objective="find peer-reviewed AI research",
+                include_domains=["arxiv.org"],
+                after_date="2026-01-01",
+                location="gb",
+                fetch_policy={"max_age_seconds": 600, "disable_cache_fallback": True},
+                client_model="claude-fable-5",
+            )
+
+            json_data = mock_post.call_args.kwargs.get("json")
+            assert json_data["objective"] == "find peer-reviewed AI research"
+            assert json_data["client_model"] == "claude-fable-5"
+
+            advanced_settings = json_data["advanced_settings"]
+            assert advanced_settings["location"] == "gb"
+            assert advanced_settings["fetch_policy"] == {
+                "max_age_seconds": 600,
+                "disable_cache_fallback": True,
+            }
+            assert advanced_settings["source_policy"]["include_domains"] == ["arxiv.org"]
+            assert advanced_settings["source_policy"]["after_date"] == "2026-01-01"
+
+            assert "include_domains" not in json_data
+            assert "after_date" not in json_data
+            assert "location" not in json_data
+            assert "fetch_policy" not in json_data
+
+    @pytest.mark.asyncio
+    async def test_response_preserves_raw_parallel_fields(self):
+        with patch(
+            "litellm.llms.custom_httpx.http_handler.AsyncHTTPHandler.post",
+            new_callable=AsyncMock,
+        ) as mock_post:
+            mock_post.return_value = _mock_response()
+
+            response = await litellm.asearch(
+                query="AI developments",
+                search_provider="parallel_ai",
+            )
+
+            dumped = response.model_dump()
+            assert dumped["search_id"] == "search_abc123"
+            assert dumped["session_id"] == "session_xyz"
+            assert dumped["parallel_usage"] == [{"name": "search_advanced", "count": 1}]
+
+            first = response.results[0].model_dump()
+            assert first["excerpts"] == ["First excerpt.", "Second excerpt."]
+
+    @pytest.mark.parametrize(
+        "mode,usage,max_results,expected_cost",
+        [
+            ("turbo", [{"name": "sku_search", "count": 1}], None, 0.001),
+            ("basic", [{"name": "sku_search", "count": 1}], None, 0.005),
+            ("advanced", [{"name": "sku_search", "count": 1}], None, 0.005),
+            (
+                "basic",
+                [
+                    {"name": "sku_search", "count": 1},
+                    {"name": "sku_search_additional_results", "count": 2},
+                ],
+                20,
+                0.007,
+            ),
+            ("basic", None, 20, 0.015),
+        ],
+    )
+    @pytest.mark.asyncio
+    async def test_search_cost_uses_mode_and_provider_usage(self, mode, usage, max_results, expected_cost, bundled_cost_map):
+        response_payload = {**MOCK_V1_RESPONSE, "usage": usage}
+        with patch(
+            "litellm.llms.custom_httpx.http_handler.AsyncHTTPHandler.post",
+            new_callable=AsyncMock,
+        ) as mock_post:
+            mock_post.return_value = _mock_response(response_payload)
+
+            response = await litellm.asearch(
+                query="AI developments",
+                search_provider="parallel_ai",
+                mode=mode,
+                max_results=max_results,
+            )
+
+            assert response._hidden_params["response_cost"] == pytest.approx(expected_cost)
+
+    @pytest.mark.asyncio
+    async def test_search_cost_treats_keyword_queries_as_one_request(self, bundled_cost_map):
+        response_payload = {
+            **MOCK_V1_RESPONSE,
+            "usage": [{"name": "sku_search", "count": 1}],
+        }
+        with patch(
+            "litellm.llms.custom_httpx.http_handler.AsyncHTTPHandler.post",
+            new_callable=AsyncMock,
+        ) as mock_post:
+            mock_post.return_value = _mock_response(response_payload)
+
+            response = await litellm.asearch(
+                query=["AI developments", "machine learning trends"],
+                search_provider="parallel_ai",
+                mode="basic",
+            )
+
+            assert response._hidden_params["response_cost"] == pytest.approx(0.005)

tests/test_litellm/llms/parallel_ai/test_parallel_ai_search_gateway.py

@@ -0,0 +1,191 @@
+"""Gateway coverage for Parallel AI Search."""
+
+from __future__ import annotations
+
+from collections.abc import Iterator
+from typing import Final
+from unittest.mock import AsyncMock
+
+import httpx
+import pytest
+from fastapi.testclient import TestClient
+
+import litellm
+from litellm import Router
+from litellm.integrations.websearch_interception.handler import (
+    WebSearchInterceptionLogger,
+)
+from litellm.llms.custom_httpx.http_handler import AsyncHTTPHandler
+from litellm.proxy import proxy_server
+from litellm.proxy._types import UserAPIKeyAuth
+from litellm.proxy.auth.user_api_key_auth import user_api_key_auth
+from litellm.types.utils import LlmProviders
+
+PARALLEL_SEARCH_URL: Final = "https://api.parallel.ai/v1/search"
+
+
+@pytest.fixture
+def client() -> TestClient:
+    return TestClient(proxy_server.app, raise_server_exceptions=False)
+
+
+@pytest.fixture
+def auth_as() -> Iterator[None]:
+    async def _authorized_request() -> UserAPIKeyAuth:
+        return UserAPIKeyAuth(
+            api_key="hashed-sk-test",
+            user_id="parallel-test-user",
+        )
+
+    previous: Final = proxy_server.app.dependency_overrides.get(user_api_key_auth)
+    proxy_server.app.dependency_overrides[user_api_key_auth] = _authorized_request
+    try:
+        yield
+    finally:
+        if previous is None:
+            proxy_server.app.dependency_overrides.pop(user_api_key_auth, None)
+        else:
+            proxy_server.app.dependency_overrides[user_api_key_auth] = previous
+
+
+def _parallel_search_body() -> dict[str, object]:
+    return {
+        "search_id": "search_parallel_gateway",
+        "results": [
+            {
+                "url": "https://example.com/parallel",
+                "title": "Parallel result",
+                "publish_date": "2026-08-13",
+                "excerpts": ["First excerpt", "Second excerpt"],
+            }
+        ],
+        "usage": [{"name": "sku_search", "count": 1}],
+    }
+
+
+def _parallel_router() -> Router:
+    return Router(
+        model_list=[],
+        search_tools=[
+            {
+                "search_tool_name": "parallel-search",
+                "litellm_params": {
+                    "search_provider": "parallel_ai",
+                    "api_key": "parallel-search-key",
+                    "mode": "turbo",
+                },
+            }
+        ],
+        num_retries=0,
+    )
+
+
+def _mock_async_post(
+    monkeypatch,
+    *,
+    url: str,
+    response_body: dict[str, object],
+) -> AsyncMock:
+    response = httpx.Response(
+        status_code=200,
+        json=response_body,
+        request=httpx.Request("POST", url),
+    )
+    mock_post = AsyncMock(return_value=response)
+    monkeypatch.setattr(AsyncHTTPHandler, "post", mock_post)
+    return mock_post
+
+
+def test_parallel_search_gateway_route(client, auth_as, monkeypatch):
+    """The named search route selects its configured Parallel Search tool.
+
+    The tool-level `mode` must survive the router hop, so the upstream request
+    is sent as `turbo` rather than falling back to the adapter default.
+    """
+    monkeypatch.setattr(litellm, "model_cost", litellm.get_model_cost_map(url=""))
+    monkeypatch.setattr(proxy_server, "llm_router", _parallel_router())
+    mock_post = _mock_async_post(
+        monkeypatch,
+        url=PARALLEL_SEARCH_URL,
+        response_body=_parallel_search_body(),
+    )
+
+    response = client.post(
+        "/v1/search/parallel-search",
+        json={"query": "Parallel AI news", "max_results": 3},
+    )
+
+    assert response.status_code == 200, response.text
+    assert response.json()["results"] == [
+        {
+            "title": "Parallel result",
+            "url": "https://example.com/parallel",
+            "snippet": "First excerpt ... Second excerpt",
+            "date": "2026-08-13",
+            "last_updated": None,
+            "excerpts": ["First excerpt", "Second excerpt"],
+        }
+    ]
+
+    request_kwargs = mock_post.await_args.kwargs
+    assert request_kwargs["url"] == PARALLEL_SEARCH_URL
+    assert request_kwargs["headers"]["x-api-key"] == "parallel-search-key"
+    assert request_kwargs["json"] == {
+        "objective": "Parallel AI news",
+        "search_queries": ["Parallel AI news"],
+        "mode": "turbo",
+        "advanced_settings": {"max_results": 3},
+    }
+
+
+@pytest.mark.asyncio
+async def test_web_search_interception_executes_parallel_search(monkeypatch):
+    """An intercepted web-search call uses the configured Parallel Search tool."""
+    monkeypatch.setattr(litellm, "model_cost", litellm.get_model_cost_map(url=""))
+    monkeypatch.setattr(proxy_server, "llm_router", _parallel_router())
+    mock_post = _mock_async_post(
+        monkeypatch,
+        url=PARALLEL_SEARCH_URL,
+        response_body=_parallel_search_body(),
+    )
+    logger = WebSearchInterceptionLogger(
+        enabled_providers=[LlmProviders.OPENAI],
+        search_tool_name="parallel-search",
+    )
+
+    plan = await logger.async_build_responses_agentic_loop_plan(
+        tools={
+            "tool_calls": [
+                {
+                    "id": "fc_parallel",
+                    "call_id": "fc_parallel",
+                    "type": "function_call",
+                    "name": "litellm_web_search",
+                    "arguments": '{"query":"Parallel AI news"}',
+                    "input": {"query": "Parallel AI news"},
+                }
+            ]
+        },
+        model="gpt-5",
+        messages=[{"role": "user", "content": "Research Parallel"}],
+        response=None,
+        optional_params={"tools": [{"type": "function", "name": "litellm_web_search"}]},
+        logging_obj=None,
+        stream=False,
+        kwargs={"custom_llm_provider": "openai"},
+    )
+
+    assert plan.run_agentic_loop is True
+    assert plan.request_patch is not None
+    assert plan.request_patch.messages[-1] == {
+        "type": "function_call_output",
+        "call_id": "fc_parallel",
+        "output": (
+            "Title: Parallel result\nURL: https://example.com/parallel\nSnippet: First excerpt ... Second excerpt"
+        ),
+    }
+
+    request_kwargs = mock_post.await_args.kwargs
+    assert request_kwargs["url"] == PARALLEL_SEARCH_URL
+    assert request_kwargs["headers"]["x-api-key"] == "parallel-search-key"
+    assert request_kwargs["json"]["mode"] == "turbo"