diff --git a/EvoScientist/llm/model_config.py b/EvoScientist/llm/model_config.py
index 0e066d8..5bea899 100644
--- a/EvoScientist/llm/model_config.py
+++ b/EvoScientist/llm/model_config.py
@@ -1385,7 +1385,7 @@ def _parse_v3_config(
                     allowed=_REASONING_EFFORTS - {"disabled"},
                 )
                 default_reasoning_effort = str(
-                    reasoning_policy.get("default_effort") or "high"
+                    reasoning_policy.get("default_effort") or "medium"
                 )
                 if default_reasoning_effort not in allowed_reasoning_efforts:
                     raise EvoRuntimeError(
diff --git a/EvoScientist/llm/models.py b/EvoScientist/llm/models.py
index 84a33b6..520db30 100644
--- a/EvoScientist/llm/models.py
+++ b/EvoScientist/llm/models.py
@@ -593,7 +593,7 @@ def get_chat_model(
         # passback (OpenRouter's `/responses` beta is stateless, store=false —
         # "Item with id 'rs_...' not found"); the patch strips them on passback,
         # so enabling `summary` is safe. See langchain-ai/langchain#37777.
-        kwargs.setdefault("reasoning", {"effort": "high", "summary": "auto"})
+        kwargs.setdefault("reasoning", {"effort": "medium", "summary": "auto"})
         # App attribution (issue #339): identify EvoScientist to OpenRouter so
         # usage is credited to the project (app rankings, model app tabs,
         # analytics) rather than langchain-openrouter's LangChain-branded
diff --git a/tests/test_llm.py b/tests/test_llm.py
index dc604bb..6fd848c 100644
--- a/tests/test_llm.py
+++ b/tests/test_llm.py
@@ -459,7 +459,7 @@ class TestThirdPartyRouting:
         call_kwargs = mock_init.call_args[1]
         assert call_kwargs["model_provider"] == "openrouter"
         assert call_kwargs["api_key"] == "or-key-456"
-        assert call_kwargs["reasoning"] == {"effort": "high", "summary": "auto"}
+        assert call_kwargs["reasoning"] == {"effort": "medium", "summary": "auto"}
 
     @patch("EvoScientist.llm.models.init_chat_model")
     def test_openrouter_reasoning_user_override(self, mock_init, monkeypatch):
@@ -488,7 +488,7 @@ class TestThirdPartyRouting:
         get_chat_model("x-ai/grok-4.3", provider="openrouter")
 
         call_kwargs = mock_init.call_args[1]
-        assert call_kwargs["reasoning"] == {"effort": "high", "summary": "auto"}
+        assert call_kwargs["reasoning"] == {"effort": "medium", "summary": "auto"}
 
     # --- OpenRouter app attribution (issue #339) ---
 
@@ -599,7 +599,7 @@ class TestThirdPartyRouting:
 
         call_kwargs = mock_init.call_args[1]
         # Existing behavior intact.
-        assert call_kwargs["reasoning"] == {"effort": "high", "summary": "auto"}
+        assert call_kwargs["reasoning"] == {"effort": "medium", "summary": "auto"}
         assert call_kwargs["model_kwargs"]["cache_control"] == {"type": "ephemeral"}
         # Attribution added alongside.
         assert call_kwargs["app_url"] == "https://github.com/EvoScientist/EvoScientist"
diff --git a/tests/test_provider_model_config_v3.py b/tests/test_provider_model_config_v3.py
index 0d07d6d..c53a0a0 100644
--- a/tests/test_provider_model_config_v3.py
+++ b/tests/test_provider_model_config_v3.py
@@ -802,6 +802,20 @@ def test_model_reasoning_policy_can_restrict_efforts_and_enable_max() -> None:
     assert parsed.reasoning_enabled_params == {"reasoning": "high"}
 
 
+def test_model_reasoning_policy_defaults_to_medium_without_explicit_effort() -> None:
+    payload = v3_payload()
+    provider = next(item for item in payload["providers"] if item["adapter_id"] == "openai")
+    model = provider["models"][0]
+    model["capabilities"]["thinking"] = True
+    # 不设置 reasoning_policy.default_effort → 兜底应为 medium
+
+    config = EvoModelConfig.parse(payload, require_evidence=False)
+    parsed = config.providers[provider["provider_id"]].models[model["model_key"]]
+
+    assert parsed.supports_reasoning is True
+    assert parsed.reasoning_enabled_params == {"reasoning": "medium"}
+
+
 def test_generic_openai_compatible_adapter_supports_standard_model_discovery() -> None:
     registration = get_adapter_registry().get(
         "generic-openai-compatible", "generic-openai-compatible-v1"