Merge pull request #38439 from BerriAI/litellm_messages_tool_usage_cost_header

fix(anthropic_adapter): carry web search cost into /v1/messages breakdown headers

basedpyright-code-budget.json

@@ -84,7 +84,7 @@
     "limit": 56
   },
   "reportPrivateUsage": {
-    "limit": 1810
+    "limit": 1808
   },
   "reportRedeclaration": {
     "limit": 8
@@ -135,7 +135,7 @@
     "limit": 21
   },
   "reportUnusedFunction": {
-    "limit": 139
+    "limit": 138
   },
   "reportUnusedImport": {
     "limit": 544

litellm/litellm_core_utils/llm_cost_calc/tool_call_cost_tracking.py

@@ -7,7 +7,7 @@ from typing import Any, Final, Literal
 
 import litellm
 from litellm.constants import OPENAI_FILE_SEARCH_COST_PER_1K_CALLS
-from litellm.litellm_core_utils.llm_cost_calc.utils import _get_web_search_requests
+from litellm.litellm_core_utils.llm_cost_calc.utils import get_web_search_requests
 from litellm.types.llms.openai import (
     FileSearchTool,
     ResponsesAPIResponse,
@@ -368,7 +368,7 @@ class StandardBuiltInToolCostTracking:
             get_anthropic_web_search_requests_from_response,
         )
 
-        if usage is not None and (_get_web_search_requests(getattr(usage, "server_tool_use", None)) is not None):
+        if usage is not None and (get_web_search_requests(getattr(usage, "server_tool_use", None)) is not None):
             return usage
         web_search_requests: Final = get_anthropic_web_search_requests_from_response(response_object)
         if web_search_requests is None:
@@ -416,7 +416,7 @@ class StandardBuiltInToolCostTracking:
                 # Anthropic Claude (direct API and Vertex AI) uses server_tool_use.web_search_requests.
                 # Without this check, Claude ModelResponse always falls through to return False
                 # and _handle_web_search_cost() is never called.
-                if hasattr(usage, "server_tool_use") and _get_web_search_requests(usage.server_tool_use) is not None:
+                if hasattr(usage, "server_tool_use") and get_web_search_requests(usage.server_tool_use) is not None:
                     return True
                 # xAI reports usage.server_side_tool_usage_details.web_search_calls; a searched
                 # answer with no url_citation annotations has no other chat-path signal
@@ -431,7 +431,7 @@ class StandardBuiltInToolCostTracking:
         elif usage is not None:
             if (
                 hasattr(usage, "server_tool_use")
-                and _get_web_search_requests(usage.server_tool_use) is not None
+                and get_web_search_requests(usage.server_tool_use) is not None
                 or (
                     hasattr(usage, "prompt_tokens_details")
                     and usage.prompt_tokens_details is not None

litellm/litellm_core_utils/llm_cost_calc/utils.py

@@ -72,7 +72,7 @@ def _get_token_detail_value(details: object, key: str) -> int | None:
     return value if isinstance(value, int) else None
 
 
-def _get_web_search_requests(server_tool_use: Any) -> int | None:
+def get_web_search_requests(server_tool_use: Any) -> int | None:
     """
     Tolerantly read ``web_search_requests`` from a ``server_tool_use`` value
     that may be ``None``, a ``dict``, a ``ServerToolUse`` pydantic instance,

litellm/llms/anthropic/cost_calculation.py

@@ -8,9 +8,9 @@ from typing import TYPE_CHECKING, Final, Optional
 from pydantic import BaseModel, ValidationError
 
 from litellm.litellm_core_utils.llm_cost_calc.utils import (
-    _get_web_search_requests,
     generic_cost_per_token,
     get_provider_specific_geo_multiplier,
+    get_web_search_requests,
 )
 
 if TYPE_CHECKING:
@@ -104,7 +104,7 @@ def get_cost_for_anthropic_web_search(
 
     if usage is None:
         return 0.0
-    web_search_requests: Final = _get_web_search_requests(getattr(usage, "server_tool_use", None))
+    web_search_requests: Final = get_web_search_requests(getattr(usage, "server_tool_use", None))
     if web_search_requests is None:
         return 0.0
 

litellm/llms/anthropic/experimental_pass_through/adapters/transformation.py

@@ -99,6 +99,7 @@ from litellm.types.llms.anthropic import (
     ContextManagementResponse,
     MessageBlockDelta,
     MessageDelta,
+    ServerToolUsage,
     StreamingContentBlockDeltaType,
     UsageDelta,
     UsageIteration,
@@ -1354,10 +1355,24 @@ class LiteLLMAnthropicMessagesAdapter:
             return explicit_value
         return cls._first_positive_prompt_tokens_detail_value(usage, ("cache_creation_tokens", "cache_write_tokens"))
 
+    @classmethod
+    def _get_web_search_request_count(cls, usage: Usage) -> int:
+        from litellm.litellm_core_utils.llm_cost_calc.utils import (
+            get_web_search_requests,
+        )
+
+        from_server_tool_use: Final = cls._positive_int(
+            get_web_search_requests(getattr(usage, "server_tool_use", None))
+        )
+        if from_server_tool_use > 0:
+            return from_server_tool_use
+        return cls._first_positive_prompt_tokens_detail_value(usage, ("web_search_requests",))
+
     @classmethod
     def _translate_openai_usage_to_anthropic_usage_delta(cls, usage: Usage) -> UsageDelta:
         cache_read_input_tokens: Final = cls._get_cache_read_input_tokens(usage)
         cache_creation_input_tokens: Final = cls._get_cache_creation_input_tokens(usage)
+        web_search_requests: Final = cls._get_web_search_request_count(usage)
         input_tokens: Final = max(
             (usage.prompt_tokens or 0) - cache_read_input_tokens - cache_creation_input_tokens,
             0,
@@ -1371,6 +1386,11 @@ class LiteLLMAnthropicMessagesAdapter:
             usage_delta["cache_creation_input_tokens"] = cache_creation_input_tokens
         if cache_read_input_tokens > 0:
             usage_delta["cache_read_input_tokens"] = cache_read_input_tokens
+        if web_search_requests > 0:
+            return UsageDelta(
+                **usage_delta,
+                server_tool_use=ServerToolUsage(web_search_requests=web_search_requests),
+            )
         return usage_delta
 
     @classmethod

litellm/llms/gemini/cost_calculator.py

@@ -39,28 +39,33 @@ def cost_per_web_search_request(usage: "Usage", model_info: "ModelInfo") -> floa
     ``model_info`` when available, falling back to $0.035 for models not
     yet updated in the pricing JSON.
     """
+    from litellm.litellm_core_utils.llm_cost_calc.utils import get_web_search_requests
     from litellm.types.utils import PromptTokensDetailsWrapper
 
     _DEFAULT_COST: Final = 35e-3
     search_costs: Final = model_info.get("search_context_cost_per_query") or {}
     _cost: Final = search_costs.get("search_context_size_medium", _DEFAULT_COST)
 
-    number_of_web_search_requests = 0
-    if (
-        usage is not None
-        and usage.prompt_tokens_details is not None
-        and isinstance(usage.prompt_tokens_details, PromptTokensDetailsWrapper)
-        and hasattr(usage.prompt_tokens_details, "web_search_requests")
-        and usage.prompt_tokens_details.web_search_requests is not None
-    ):
-        number_of_web_search_requests = usage.prompt_tokens_details.web_search_requests
+    requests_from_prompt_details: Final = (
+        usage.prompt_tokens_details.web_search_requests
+        if (
+            usage is not None
+            and usage.prompt_tokens_details is not None
+            and isinstance(usage.prompt_tokens_details, PromptTokensDetailsWrapper)
+            and hasattr(usage.prompt_tokens_details, "web_search_requests")
+            and usage.prompt_tokens_details.web_search_requests is not None
+        )
+        else None
+    )
+    requests_from_server_tool_use: Final = get_web_search_requests(getattr(usage, "server_tool_use", None))
+    number_of_web_search_requests: Final = requests_from_prompt_details or requests_from_server_tool_use or 0
 
-    # per_prompt billing: clamp to 1 (flat fee per grounded API call)
     billing_mode: Final = model_info.get("web_search_billing_unit") or "per_prompt"
-    if number_of_web_search_requests > 0 and billing_mode == "per_prompt":
-        number_of_web_search_requests = 1
+    billable_requests: Final = (
+        1 if (number_of_web_search_requests > 0 and billing_mode == "per_prompt") else number_of_web_search_requests
+    )
 
-    return _cost * number_of_web_search_requests
+    return _cost * billable_requests
 
 
 GOOGLE_MAPS_GROUNDING_DEFAULT_COST_PER_QUERY: Final = 14e-3

litellm/types/llms/anthropic.py

@@ -502,11 +502,16 @@ class MessageDelta(TypedDict, total=False):
     stop_reason: str | None
 
 
+class ServerToolUsage(TypedDict, total=False):
+    web_search_requests: ReadOnly[int]
+
+
 class UsageDelta(TypedDict, total=False):
     input_tokens: int
     output_tokens: int
     cache_creation_input_tokens: int
     cache_read_input_tokens: int
+    server_tool_use: ReadOnly[ServerToolUsage]
 
 
 class AppliedEdit(TypedDict, total=False):

litellm/types/llms/anthropic_messages/anthropic_response.py

@@ -1,11 +1,12 @@
 from typing import Any, Literal, TypeAlias
 
-from typing_extensions import NotRequired, TypedDict
+from typing_extensions import NotRequired, ReadOnly, TypedDict
 
 from litellm.types.llms.anthropic import (
     AnthropicResponseContentBlockText,
     AnthropicResponseContentBlockToolUse,
     ContextManagementResponse,
+    ServerToolUsage,
 )
 
 
@@ -71,6 +72,11 @@ class AnthropicUsage(TypedDict, total=False):
     cache_creation_input_tokens: int
     cache_read_input_tokens: int
 
+    """
+    Server-side tool usage (e.g. web search request counts)
+    """
+    server_tool_use: NotRequired[ReadOnly[ServerToolUsage]]
+
 
 class AnthropicMessagesResponse(TypedDict, total=False):
     """

tests/test_litellm/litellm_core_utils/llm_cost_calc/test_tool_call_cost_tracking_dict_safety.py

@@ -8,10 +8,9 @@ See https://github.com/BerriAI/litellm/issues/26153.
 
 import pytest
 
-
 from litellm.litellm_core_utils.llm_cost_calc.tool_call_cost_tracking import (
     StandardBuiltInToolCostTracking,
-    _get_web_search_requests,
+    get_web_search_requests,
 )
 from litellm.types.utils import ModelResponse, ServerToolUse, Usage
 
@@ -28,25 +27,25 @@ class _UsageWithDictServerToolUse:
 
 
 def test_get_web_search_requests_handles_none():
-    assert _get_web_search_requests(None) is None
+    assert get_web_search_requests(None) is None
 
 
 def test_get_web_search_requests_handles_dict():
-    assert _get_web_search_requests({"web_search_requests": 5}) == 5
+    assert get_web_search_requests({"web_search_requests": 5}) == 5
 
 
 def test_get_web_search_requests_handles_dict_missing_key():
-    assert _get_web_search_requests({}) is None
+    assert get_web_search_requests({}) is None
 
 
 def test_get_web_search_requests_handles_pydantic():
     stu = ServerToolUse(web_search_requests=7)
-    assert _get_web_search_requests(stu) == 7
+    assert get_web_search_requests(stu) == 7
 
 
 def test_get_web_search_requests_handles_pydantic_with_none_value():
     stu = ServerToolUse()
-    assert _get_web_search_requests(stu) is None
+    assert get_web_search_requests(stu) is None
 
 
 def test_response_object_includes_web_search_call_with_dict_server_tool_use():

tests/test_litellm/llms/anthropic/experimental_pass_through/adapters/test_anthropic_experimental_pass_through_adapters_transformation.py

@@ -3997,3 +3997,98 @@ def test_translate_anthropic_messages_to_openai_carries_midturn_system_prompt_ca
     assert result == [
         {"role": "system", "content": [{"type": "text", "text": "fix", "prompt_cache_breakpoint": explicit}]}
     ]
+
+
+def _openai_response_with_usage(usage: Usage) -> ModelResponse:
+    return ModelResponse(
+        id="resp_web_search",
+        model="gemini-3-flash-preview",
+        choices=[
+            Choices(
+                finish_reason="stop",
+                index=0,
+                message=Message(role="assistant", content="searched"),
+            )
+        ],
+        usage=usage,
+    )
+
+
+def test_translate_openai_response_to_anthropic_maps_gemini_web_search_usage():
+    from litellm.types.utils import PromptTokensDetailsWrapper
+
+    usage = Usage(
+        prompt_tokens=385,
+        completion_tokens=566,
+        total_tokens=951,
+        prompt_tokens_details=PromptTokensDetailsWrapper(text_tokens=385, web_search_requests=2),
+    )
+
+    anthropic_response = LiteLLMAnthropicMessagesAdapter().translate_openai_response_to_anthropic(
+        response=_openai_response_with_usage(usage)
+    )
+
+    assert anthropic_response["usage"]["server_tool_use"] == {"web_search_requests": 2}
+
+
+def test_translate_openai_response_to_anthropic_maps_server_tool_use_web_search_usage():
+    from litellm.types.utils import ServerToolUse
+
+    usage = Usage(
+        prompt_tokens=100,
+        completion_tokens=40,
+        total_tokens=140,
+        server_tool_use=ServerToolUse(web_search_requests=3),
+    )
+
+    anthropic_response = LiteLLMAnthropicMessagesAdapter().translate_openai_response_to_anthropic(
+        response=_openai_response_with_usage(usage)
+    )
+
+    assert anthropic_response["usage"]["server_tool_use"] == {"web_search_requests": 3}
+
+
+def test_translate_openai_response_to_anthropic_omits_server_tool_use_without_web_search():
+    usage = Usage(prompt_tokens=100, completion_tokens=40, total_tokens=140)
+
+    anthropic_response = LiteLLMAnthropicMessagesAdapter().translate_openai_response_to_anthropic(
+        response=_openai_response_with_usage(usage)
+    )
+
+    assert "server_tool_use" not in anthropic_response["usage"]
+
+
+def test_completion_cost_on_translated_anthropic_response_includes_web_search():
+    from litellm.types.utils import PromptTokensDetailsWrapper
+
+    adapter = LiteLLMAnthropicMessagesAdapter()
+    with_search = adapter.translate_openai_response_to_anthropic(
+        response=_openai_response_with_usage(
+            Usage(
+                prompt_tokens=385,
+                completion_tokens=566,
+                total_tokens=951,
+                prompt_tokens_details=PromptTokensDetailsWrapper(text_tokens=385, web_search_requests=2),
+            )
+        )
+    )
+    without_search = adapter.translate_openai_response_to_anthropic(
+        response=_openai_response_with_usage(Usage(prompt_tokens=385, completion_tokens=566, total_tokens=951))
+    )
+
+    cost_with_search = litellm.completion_cost(
+        completion_response=with_search,
+        model="gemini/gemini-3-flash-preview",
+        call_type="anthropic_messages",
+    )
+    cost_without_search = litellm.completion_cost(
+        completion_response=without_search,
+        model="gemini/gemini-3-flash-preview",
+        call_type="anthropic_messages",
+    )
+
+    per_query_cost = litellm.model_cost["gemini/gemini-3-flash-preview"]["search_context_cost_per_query"][
+        "search_context_size_medium"
+    ]
+    assert per_query_cost > 0
+    assert cost_with_search - cost_without_search == pytest.approx(2 * per_query_cost)

tests/test_litellm/llms/anthropic/test_cost_calculation_dict_safety.py

@@ -8,10 +8,9 @@ See https://github.com/BerriAI/litellm/issues/26153.
 
 import pytest
 
-
 from litellm.llms.anthropic.cost_calculation import (
-    _get_web_search_requests,
     get_cost_for_anthropic_web_search,
+    get_web_search_requests,
 )
 from litellm.types.utils import ModelInfo, ServerToolUse
 
@@ -33,19 +32,19 @@ def _make_model_info(cost_per_query: float = 0.01) -> ModelInfo:
 
 
 def test_get_web_search_requests_handles_none():
-    assert _get_web_search_requests(None) is None
+    assert get_web_search_requests(None) is None
 
 
 def test_get_web_search_requests_handles_dict():
-    assert _get_web_search_requests({"web_search_requests": 4}) == 4
+    assert get_web_search_requests({"web_search_requests": 4}) == 4
 
 
 def test_get_web_search_requests_handles_dict_missing_key():
-    assert _get_web_search_requests({}) is None
+    assert get_web_search_requests({}) is None
 
 
 def test_get_web_search_requests_handles_pydantic():
-    assert _get_web_search_requests(ServerToolUse(web_search_requests=2)) == 2
+    assert get_web_search_requests(ServerToolUse(web_search_requests=2)) == 2
 
 
 def test_get_cost_for_anthropic_web_search_with_dict_server_tool_use():

tests/test_litellm/llms/gemini/test_cost_calculator.py

@@ -84,6 +84,65 @@ def test_no_usage_details():
     assert cost == 0.0
 
 
+def _make_server_tool_use_usage(web_search_requests: int) -> Usage:
+    from litellm.types.utils import ServerToolUse
+
+    return Usage(
+        prompt_tokens=100,
+        completion_tokens=50,
+        total_tokens=150,
+        server_tool_use=ServerToolUse(web_search_requests=web_search_requests),
+    )
+
+
+def test_server_tool_use_fallback_per_query_billing():
+    """Usage reconstructed from an Anthropic-format response carries the count in
+    server_tool_use, not prompt_tokens_details; per_query billing prices each request."""
+    model_info = {
+        "key": "gemini/gemini-3-flash-preview",
+        "web_search_billing_unit": "per_query",
+        "search_context_cost_per_query": {
+            "search_context_size_medium": 0.014,
+        },
+    }
+    cost = cost_per_web_search_request(usage=_make_server_tool_use_usage(3), model_info=model_info)
+    assert cost == pytest.approx(0.014 * 3)
+
+
+def test_server_tool_use_fallback_per_prompt_clamps_to_one():
+    """per_prompt billing clamps the server_tool_use count to one grounded prompt."""
+    model_info = {
+        "key": "gemini/gemini-2.5-flash",
+        "search_context_cost_per_query": {
+            "search_context_size_medium": 0.035,
+        },
+    }
+    cost = cost_per_web_search_request(usage=_make_server_tool_use_usage(4), model_info=model_info)
+    assert cost == pytest.approx(0.035 * 1)
+
+
+def test_prompt_tokens_details_take_precedence_over_server_tool_use():
+    """The native Gemini field wins when both counts are present."""
+    from litellm.types.utils import ServerToolUse
+
+    model_info = {
+        "key": "gemini/gemini-3-flash-preview",
+        "web_search_billing_unit": "per_query",
+        "search_context_cost_per_query": {
+            "search_context_size_medium": 0.014,
+        },
+    }
+    usage = Usage(
+        prompt_tokens=100,
+        completion_tokens=50,
+        total_tokens=150,
+        prompt_tokens_details=PromptTokensDetailsWrapper(web_search_requests=2),
+        server_tool_use=ServerToolUse(web_search_requests=5),
+    )
+    cost = cost_per_web_search_request(usage=usage, model_info=model_info)
+    assert cost == pytest.approx(0.014 * 2)
+
+
 def _make_maps_usage(google_maps_grounding_requests: int) -> Usage:
     return Usage(
         prompt_tokens=100,

type-discipline-budget.json

@@ -27,7 +27,7 @@
     "limit": 0
   },
   "LIT010": {
-    "limit": 16619
+    "limit": 16616
   },
   "LIT011": {
     "limit": 5583