diff --git a/CHANGELOG.md b/CHANGELOG.md
index 2aff2ba..2ff6bd1 100644
--- a/CHANGELOG.md
+++ b/CHANGELOG.md
@@ -8,6 +8,7 @@
 - Renamed the active call-budget, fallback, run-log, and telemetry fields from OpenRouter-specific names to AI Gateway names.
 - AI Gateway requests rely on the existing JSON-only prompt and strict parser instead of `response_format=json_object`, which DeepSeek V4 rejects through the Gateway.
 - Added an explicit 32,768-token output ceiling and low-cost fallback recovery for HTTP 200 responses whose model content is truncated or otherwise invalid JSON.
+- Normalize pure Markdown JSON fences before strict parsing so a valid fenced response does not spend a fallback call.
 - A zero AI Gateway call budget now completes as a real dry run instead of failing the empty-update guard.
 
 ## v0.19.5 - 2026-07-18
diff --git a/scripts/cloud_agent_runner.py b/scripts/cloud_agent_runner.py
index aa021e3..e689fd3 100644
--- a/scripts/cloud_agent_runner.py
+++ b/scripts/cloud_agent_runner.py
@@ -1296,11 +1296,16 @@ def screening_schema_text(root: Path | None = None) -> str:
     )
 
 
-def parse_screening_json(text: str) -> dict[str, Any]:
+def normalize_model_json_text(text: str) -> str:
     cleaned = text.strip()
     if cleaned.startswith("```"):
-        cleaned = re.sub(r"^```(?:json)?\s*", "", cleaned)
+        cleaned = re.sub(r"^```(?:json)?\s*", "", cleaned, flags=re.IGNORECASE)
         cleaned = re.sub(r"\s*```$", "", cleaned)
+    return cleaned.strip()
+
+
+def parse_screening_json(text: str) -> dict[str, Any]:
+    cleaned = normalize_model_json_text(text)
     try:
         data = json.loads(cleaned)
     except json.JSONDecodeError:
@@ -5523,7 +5528,7 @@ def call_ai_gateway_model(prompt: str, model: str) -> dict[str, Any]:
         if isinstance(parsed, dict) and parsed.get("error") and "choices" not in parsed:
             last_error = f"Vercel AI Gateway error envelope for {candidate_model}: {str(parsed.get('error'))[:300]}"
             continue
-        content = response_output_text(parsed).strip()
+        content = normalize_model_json_text(response_output_text(parsed))
         if not content:
             # A 200 with empty content (provider hiccup): retry the fallback
             # chain instead of failing later with "Model did not return valid JSON: ".
@@ -5544,6 +5549,15 @@ def call_ai_gateway_model(prompt: str, model: str) -> dict[str, Any]:
                 f" (finish_reason={finish_reason or 'unknown'}): {content[:300]}"
             )
             continue
+        # Models sometimes wrap an otherwise valid JSON object in a Markdown
+        # fence despite the system instruction. Normalize it once here so all
+        # downstream strict parsers receive the same valid JSON text without
+        # spending another fallback call.
+        choices = parsed.get("choices") if isinstance(parsed, dict) else None
+        if isinstance(choices, list) and choices and isinstance(choices[0], dict):
+            message = choices[0].get("message")
+            if isinstance(message, dict):
+                message["content"] = content
         return parsed
     raise SystemExit(last_error or "Vercel AI Gateway API error.")
 
diff --git a/tests/test_cloud_agent_runner.py b/tests/test_cloud_agent_runner.py
index 38eb950..db8b5de 100644
--- a/tests/test_cloud_agent_runner.py
+++ b/tests/test_cloud_agent_runner.py
@@ -2469,6 +2469,28 @@ def test_ai_gateway_call_falls_back_on_invalid_json_content(self) -> None:
         self.assertEqual(json.loads(cloud_agent_runner.response_output_text(parsed))["summary"], "ok")
         self.assertIn("model-a->model-b", cloud_agent_runner.RUN_AUDIT["fallbacks"])
 
+    def test_ai_gateway_accepts_fenced_json_without_fallback(self) -> None:
+        response = mock.MagicMock()
+        response.__enter__.return_value.read.return_value = json.dumps(
+            {
+                "choices": [
+                    {
+                        "message": {"content": '```json\n{"summary": "ok"}\n```'},
+                        "finish_reason": "stop",
+                    }
+                ]
+            }
+        ).encode("utf-8")
+        with mock.patch.dict(
+            os.environ,
+            {"AI_GATEWAY_API_KEY": "x", "AI_GATEWAY_FALLBACK_MODELS": "model-b"},
+            clear=False,
+        ):
+            with mock.patch.object(urllib.request, "urlopen", return_value=response) as urlopen_mock:
+                parsed = cloud_agent_runner.call_ai_gateway_model("prompt", "model-a")
+        self.assertEqual(json.loads(cloud_agent_runner.response_output_text(parsed))["summary"], "ok")
+        urlopen_mock.assert_called_once()
+
     def test_max_response_chars_generous_for_all_tasks(self) -> None:
         # Issue #59 round 3: the daily legitimately produced 75.3k chars under
         # the v0.19 funnel; every task now gets the 96k cap by default.