diff --git a/packages/llm_gateway/tests/test_new_adapters.py b/packages/llm_gateway/tests/test_new_adapters.py
index 04fd623..ad15d03 100644
--- a/packages/llm_gateway/tests/test_new_adapters.py
+++ b/packages/llm_gateway/tests/test_new_adapters.py
@@ -164,6 +164,51 @@ async def test_gemini_structured_output() -> None:
     assert "config" in models.last_kwargs
 
 
+async def test_anthropic_structured_output_threads_usage() -> None:
+    # 结构化输出必须记真实 usage（CR-H3）：instructor 的 create_with_completion 回
+    # (parsed, raw)，raw.usage 带真实 token 数；网关据此记账（§4.8）。
+    class _RawUsage:
+        input_tokens = 120
+        output_tokens = 80
+        cache_read_input_tokens = 30
+
+    class _Raw:
+        usage = _RawUsage()
+
+    class _FakeStructured:
+        def __init__(self) -> None:
+            self.last_kwargs: dict[str, Any] = {}
+
+        async def create_with_completion(
+            self, *, response_model: type[BaseModel], **kwargs: Any
+        ) -> tuple[BaseModel, Any]:
+            self.last_kwargs = kwargs
+            return response_model(score=7), _Raw()
+
+        # 修前路径调 `.create`（仅回 parsed）——保留它，令 RED 是值不符（0!=120）
+        # 而非 AttributeError，坐实旧代码硬编码零 usage 的缺陷。
+        async def create(self, *, response_model: type[BaseModel], **kwargs: Any) -> BaseModel:
+            return response_model(score=7)
+
+    fake = _FakeStructured()
+    client = _FakeAnthropicClient(_FakeAnthropicMessages("", 0, 0, 0))
+    adapter = AnthropicAdapter("anthropic", client, structured_client=fake)
+    req = LlmRequest(
+        tier="analyst",
+        input="打分",
+        output_schema=Out,
+        scope=Scope(user_id=uuid.UUID(int=1)),
+    )
+
+    result = await adapter.complete(req, "claude-3-5-sonnet")
+
+    assert isinstance(result.parsed, Out)
+    assert result.parsed.score == 7
+    assert result.usage.input_tokens == 120
+    assert result.usage.output_tokens == 80
+    assert result.usage.cache_read_tokens == 30
+
+
 async def test_anthropic_stream_yields_text_then_usage() -> None:
     class _StreamEvent:
         def __init__(self, **kw: Any) -> None:
diff --git a/packages/llm_gateway/ww_llm_gateway/adapters/anthropic.py b/packages/llm_gateway/ww_llm_gateway/adapters/anthropic.py
index 5064b0a..5d12ede 100644
--- a/packages/llm_gateway/ww_llm_gateway/adapters/anthropic.py
+++ b/packages/llm_gateway/ww_llm_gateway/adapters/anthropic.py
@@ -46,7 +46,9 @@ class AnthropicClient(Protocol):
 
 
 class StructuredAnthropic(Protocol):
-    async def create(self, *, response_model: type[BaseModel], **kwargs: Any) -> BaseModel: ...
+    async def create_with_completion(
+        self, *, response_model: type[BaseModel], **kwargs: Any
+    ) -> tuple[BaseModel, Any]: ...
 
 
 def _is_transient(exc: Exception) -> bool:
@@ -115,7 +117,7 @@ class AnthropicAdapter:
             # 注入客户端是 `AnthropicClient` Protocol（保留测试替身缝），
             # 但 instructor.from_anthropic 只重载真实 SDK 的具体类型；
             # 本适配器异步 → 转 `AsyncAnthropic` 选中 AsyncInstructor 重载，
-            # 返回的 AsyncInstructor 鸭子匹配 StructuredAnthropic。
+            # 返回的 AsyncInstructor 鸭子匹配 StructuredAnthropic（含 create_with_completion）。
             async_client = cast("AsyncAnthropic", self._client)
             self._structured_client = cast(
                 "StructuredAnthropic", instructor.from_anthropic(async_client)
@@ -156,9 +158,11 @@ class AnthropicAdapter:
         system = _system_blocks(req)
         if system:
             kwargs["system"] = system
-        parsed = await self._structured().create(response_model=req.output_schema, **kwargs)
-        # instructor.from_anthropic 默认不回 raw usage（按名隐藏）；usage 经流/text 路径覆盖。
-        usage = ProviderUsage(input_tokens=0, output_tokens=0)
+        parsed, raw = await self._structured().create_with_completion(
+            response_model=req.output_schema, **kwargs
+        )
+        # 从 raw completion 提取真实 usage（记账真源，§4.8）；不再硬编码零（CR-H3）。
+        usage = _usage_from(getattr(raw, "usage", None))
         return ProviderResult(text=parsed.model_dump_json(), usage=usage, parsed=parsed)
 
     async def stream(self, req: LlmRequest, model: str) -> AsyncIterator[StreamChunk]: