diff --git a/backend/app/data/catalog.py b/backend/app/data/catalog.py
index 45eec1bb..058c8b16 100644
--- a/backend/app/data/catalog.py
+++ b/backend/app/data/catalog.py
@@ -1570,6 +1570,8 @@ async def resolve_draft_model(session: AsyncSession) -> tuple[Any, str]:
 
     不关思考（spec 里不传 thinking，按接入默认）、温度 0、额度 MODEL_MAX_TOKENS，服务方嫌大时退回
     MODEL_FALLBACK_MAX_TOKENS（DraftModel）。单次调用的时限 MODEL_TIMEOUT_S 也交给 SDK，免得它先按全局的 5 分钟断开。
+    SDK 自己不重试（max_retries=0）：429、5xx、网络错误、超时由起草引擎按失败原因重试（最多 MODEL_RETRIES 次），
+    不然两层叠起来一段最多发 9 次请求，报错却写「已重试 2 次」，网关正在限流时还会放大流量。
     只建模型对象，不发请求。
 
     和 recipe_ai.resolve_draft_model 一样不用 get_chat_model：它在没有接入时静默退回演示模型，拿演示模型
@@ -1585,7 +1587,7 @@ async def resolve_draft_model(session: AsyncSession) -> tuple[Any, str]:
     except DraftModelUnavailable as e:
         raise CatalogModelUnavailable(str(e)) from e
     spec = ModelSpec(provider=choice.provider, model=choice.model, max_tokens=MODEL_MAX_TOKENS, temperature=0,
-                     timeout=MODEL_TIMEOUT_S)
+                     timeout=MODEL_TIMEOUT_S, max_retries=0)
     try:
         big = build_chat_model(provider, spec)
         small = build_chat_model(provider, spec.model_copy(update={"max_tokens": MODEL_FALLBACK_MAX_TOKENS}))
diff --git a/backend/app/data/catalog_draft.py b/backend/app/data/catalog_draft.py
index 8833add9..4406c5c7 100644
--- a/backend/app/data/catalog_draft.py
+++ b/backend/app/data/catalog_draft.py
@@ -23,7 +23,8 @@
 - 结构化输出这个接口不支持（400 / 422 提到 response_format、json_schema、tool……，或者模型类不支持）：这次起草
   后面改走纯文本、从正文取 JSON（纯文本那条路另外说清 JSON 的格式）。
 - 服务方嫌额度太大：退回 MODEL_FALLBACK_MAX_TOKENS 重来，这次起草后面都用它（catalog.DraftModel.shrink）。
-- 网络错误、429、5xx、超时：等一会儿重试（最多 MODEL_RETRIES 次，退避）。单次调用的时限是 MODEL_TIMEOUT_S。
+- 网络错误、429、5xx、超时：等一会儿重试（最多 MODEL_RETRIES 次，退避）。单次调用的时限是 MODEL_TIMEOUT_S。SDK
+  自己不再重试（catalog.resolve_draft_model 建模型时 max_retries=0），实际重试次数和报错写的一致。
 - 别的错误：这一段没写出来，原因照实写。
 
 **并发。** 全进程最多 MODEL_CONCURRENCY 个调用同时进行（model_gate：按事件循环各一个信号量，所有起草任务和同步
diff --git a/backend/app/providers/factory.py b/backend/app/providers/factory.py
index b94ae6d6..e7fbb29b 100644
--- a/backend/app/providers/factory.py
+++ b/backend/app/providers/factory.py
@@ -28,6 +28,9 @@ class ModelSpec(BaseModel):
     thinking: str | None = None  # off | adaptive
     effort: str | None = None  # low | medium | high | xhigh | max
     timeout: int | None = None
+    # SDK 自己遇到 429、5xx、连接错误、超时时重试几次；None 照 SDK 默认（openai、anthropic 都是 2 次）。
+    # 自己按失败原因重试的调用方（数据目录起草）设 0，免得两层重试叠在一起、实际次数和报错写的对不上
+    max_retries: int | None = None
     extra: dict[str, Any] = Field(default_factory=dict)
 
 
@@ -227,6 +230,8 @@ def build_chat_model(provider: Provider, spec: ModelSpec) -> BaseChatModel:
     # 以前只有节点显式配了才传，而没有任何地方给 spec.timeout 赋过值——所有调用
     # 都走 SDK 默认的 10 分钟
     kwargs["timeout"] = spec.timeout or settings.model_timeout_seconds
+    if spec.max_retries is not None:
+        kwargs["max_retries"] = spec.max_retries
 
     # 采样参数：能力允许才传
     if spec.temperature is not None and catalog.supports_sampling(model):
@@ -314,7 +319,8 @@ def build_chat_model(provider: Provider, spec: ModelSpec) -> BaseChatModel:
             # 且 api_key=None 时不会发出 x-api-key。
             import anthropic
 
-            client_args = {"base_url": base_url, "api_key": None, "auth_token": api_key}
+            client_args = {"base_url": base_url, "api_key": None, "auth_token": api_key,
+                           "max_retries": chat.max_retries}
             chat._client = anthropic.Anthropic(**client_args)
             chat._async_client = anthropic.AsyncAnthropic(**client_args)
         return chat
diff --git a/backend/tests/test_catalog_draft_settings.py b/backend/tests/test_catalog_draft_settings.py
index c4cbbec4..d36e73e1 100644
--- a/backend/tests/test_catalog_draft_settings.py
+++ b/backend/tests/test_catalog_draft_settings.py
@@ -196,9 +196,31 @@ async def test_catalog_draft_uses_the_draft_setting_keeps_thinking_and_a_big_bud
     assert catalog.MODEL_MAX_TOKENS == 32768 and catalog.MODEL_FALLBACK_MAX_TOKENS == 8192
     assert all(s.thinking is None and s.temperature == 0 for _, s in specs), "不关思考，按接入默认"
     assert all(s.timeout == catalog.MODEL_TIMEOUT_S == 420 for _, s in specs)
+    assert all(s.max_retries == 0 for _, s in specs), "重试只由起草引擎按失败原因来做，SDK 自己不再重试"
     assert model.budget == 32768 and model.model_name == "模型甲"
 
 
+@pytest.mark.parametrize("kind,extra", [
+    ("openai_compatible", {}),
+    ("anthropic", {}),
+    ("anthropic", {"auth_style": "bearer"}),
+], ids=["openai-compatible", "anthropic", "anthropic-bearer"])
+def test_the_built_model_does_not_retry_on_its_own_when_asked(kind, extra):
+    """ModelSpec.max_retries 一直传到 SDK 客户端（Bearer 鉴权换上的客户端也是）；不写就照 SDK 默认。"""
+    from app.core.crypto import encrypt
+    from app.providers.factory import ModelSpec, build_chat_model
+
+    provider = Provider(name="虚构接入", kind=kind, api_key=encrypt("sk-test"), base_url="https://gateway.example/v1",
+                        models=[], default_model="model-x", extra=extra)
+
+    def client_retries(chat: Any) -> int:
+        client = chat.root_async_client if kind == "openai_compatible" else chat._async_client
+        return client.max_retries
+
+    assert client_retries(build_chat_model(provider, ModelSpec(model="model-x", max_retries=0))) == 0
+    assert client_retries(build_chat_model(provider, ModelSpec(model="model-x"))) == 2
+
+
 async def test_catalog_draft_falls_back_to_the_assistant_model(mem, no_env_keys, monkeypatch):
     await _setup(mem, copilot={"provider": "", "model": "模型乙"})
     specs = _spy(monkeypatch)
@@ -213,6 +235,7 @@ async def test_recipe_ai_draft_follows_the_same_order_but_keeps_its_own_spec(mem
     assert avail.available and model_id == "模型丙" and avail.provider == "丙接入"
     assert [s.max_tokens for _, s in specs] == [A.AI_MAX_TOKENS, A.AI_FALLBACK_MAX_TOKENS]
     assert all(s.thinking == "off" and s.temperature == 0 for _, s in specs), "配方 AI 起草的思考设置不动"
+    assert all(s.max_retries is None for _, s in specs), "配方 AI 起草的重试照旧（SDK 默认）"
 
 
 async def test_a_broken_draft_setting_says_where_to_fix_it(mem, no_env_keys):