diff --git a/apps/api/src/api/v1/webhooks.py b/apps/api/src/api/v1/webhooks.py
index c73255a0e..2ad9887e3 100644
--- a/apps/api/src/api/v1/webhooks.py
+++ b/apps/api/src/api/v1/webhooks.py
@@ -159,6 +159,9 @@ async def _push_to_telegram_background(
     signoz_rps: float = 0.0,
     signoz_rps_trend: str = "stable",
     signoz_error_rate: float = 0.0,
+    # 2026-03-29 ogt: AI Token/Cost 追蹤
+    ai_tokens: int = 0,
+    ai_cost: float = 0.0,
     signoz_p99_latency: float = 0.0,
     signoz_latency_trend: str = "stable",
     signoz_trace_url: str = "",
@@ -206,6 +209,9 @@ async def _push_to_telegram_background(
             signoz_latency_trend=signoz_latency_trend,
             signoz_trace_url=signoz_trace_url,
             auto_tuning_command=auto_tuning_command,
+            # 2026-03-29 ogt: AI Token/Cost 追蹤
+            ai_tokens=ai_tokens,
+            ai_cost=ai_cost,
         )
 
         logger.info(
@@ -216,6 +222,8 @@ async def _push_to_telegram_background(
             primary_responsibility=primary_responsibility,
             confidence=confidence,
             signoz_integrated=signoz_rps > 0 or signoz_error_rate > 0,
+            ai_tokens=ai_tokens,
+            ai_cost=f"${ai_cost:.6f}",
         )
 
     except TelegramGatewayError as e:
@@ -887,8 +895,9 @@ async def receive_alert(
         }
 
         # 呼叫 OpenClaw LLM 分析 (v7.0 含 SignOz 整合)
+        # 2026-03-29 ogt: 加入 Token/Cost 追蹤
         openclaw = get_openclaw()
-        analysis_result, ai_provider, raw_response, signoz_metrics, signoz_trace_url = await openclaw.analyze_alert(alert_context)
+        analysis_result, ai_provider, raw_response, signoz_metrics, signoz_trace_url, ai_tokens, ai_cost = await openclaw.analyze_alert(alert_context)
 
         if analysis_result:
             # LLM 分析成功
@@ -1017,6 +1026,9 @@ async def receive_alert(
             signoz_latency_trend=signoz_latency_trend,
             signoz_trace_url=signoz_trace_url,
             auto_tuning_command=auto_tuning_cmd,
+            # 2026-03-29 ogt: AI Token/Cost 追蹤
+            ai_tokens=ai_tokens,
+            ai_cost=ai_cost,
         )
 
         return AlertResponse(
@@ -1251,8 +1263,9 @@ async def alertmanager_webhook(
             "labels": alert.labels,
         }
 
+        # 2026-03-29 ogt: 加入 Token/Cost 追蹤
         openclaw = get_openclaw()
-        analysis_result, ai_provider, raw_response, signoz_metrics, signoz_trace_url = await openclaw.analyze_alert(alert_context)
+        analysis_result, ai_provider, raw_response, signoz_metrics, signoz_trace_url, ai_tokens, ai_cost = await openclaw.analyze_alert(alert_context)
 
         if analysis_result:
             # analysis_result 是 OpenClawDecision Pydantic 模型
@@ -1335,6 +1348,9 @@ async def alertmanager_webhook(
                 signoz_p99_latency=signoz_metrics.p99_latency_ms if signoz_metrics else 0,
                 signoz_latency_trend=signoz_metrics.latency_trend if signoz_metrics else "stable",
                 signoz_trace_url=signoz_trace_url or "",
+                # 2026-03-29 ogt: AI Token/Cost 追蹤
+                ai_tokens=ai_tokens,
+                ai_cost=ai_cost,
             )
 
             return AlertResponse(
diff --git a/apps/api/src/services/openclaw.py b/apps/api/src/services/openclaw.py
index bda9b3be8..a022383a1 100644
--- a/apps/api/src/services/openclaw.py
+++ b/apps/api/src/services/openclaw.py
@@ -320,12 +320,21 @@ class OpenClawService:
             )
             return str(e), False
 
-    async def _call_gemini(self, prompt: str) -> tuple[str, bool]:
+    async def _call_gemini(self, prompt: str) -> tuple[str, bool, int, float]:
         """
         呼叫 Google Gemini (支援 JSON Mode)
+
+        Returns:
+            tuple: (response_text, success, total_tokens, cost_usd)
+            - response_text: LLM 回應文本
+            - success: 是否成功
+            - total_tokens: 使用的 Token 總數
+            - cost_usd: 預估成本 (USD)
+
+        2026-03-29 ogt: 加入 Token/Cost 追蹤
         """
         if not settings.GEMINI_API_KEY:
-            return "GEMINI_API_KEY not configured", False
+            return "GEMINI_API_KEY not configured", False, 0, 0.0
 
         try:
             client = await self._get_client()
@@ -350,12 +359,29 @@ class OpenClawService:
             data = response.json()
             text = data["candidates"][0]["content"]["parts"][0]["text"]
 
-            logger.info("gemini_response_received", response_length=len(text))
-            return text, True
+            # 2026-03-29 ogt: 擷取 Token 使用量
+            usage_metadata = data.get("usageMetadata", {})
+            prompt_tokens = usage_metadata.get("promptTokenCount", 0)
+            completion_tokens = usage_metadata.get("candidatesTokenCount", 0)
+            total_tokens = usage_metadata.get("totalTokenCount", prompt_tokens + completion_tokens)
+
+            # Gemini 1.5 Flash 定價 (per 1M tokens)
+            # Input: $0.075 / 1M, Output: $0.30 / 1M
+            cost_usd = (prompt_tokens * 0.000000075) + (completion_tokens * 0.0000003)
+
+            logger.info(
+                "gemini_response_received",
+                response_length=len(text),
+                prompt_tokens=prompt_tokens,
+                completion_tokens=completion_tokens,
+                total_tokens=total_tokens,
+                cost_usd=f"${cost_usd:.6f}",
+            )
+            return text, True, total_tokens, cost_usd
 
         except Exception as e:
             logger.warning("gemini_call_failed", error=str(e))
-            return str(e), False
+            return str(e), False, 0, 0.0
 
     async def _call_claude(self, prompt: str) -> tuple[str, bool]:
         """
@@ -672,7 +698,7 @@ class OpenClawService:
         alert_context: dict | None = None,
         signoz_metrics: GoldMetrics | None = None,
         cache_ttl: int = 3600,  # 1 hour default
-    ) -> tuple[str, str, bool, bool]:
+    ) -> tuple[str, str, bool, bool, int, float]:
         """
         帶快取的 LLM 呼叫包裝器
 
@@ -685,7 +711,9 @@ class OpenClawService:
             cache_ttl: 快取存活時間 (秒)
 
         Returns:
-            (response, provider, success, from_cache)
+            (response, provider, success, from_cache, total_tokens, cost_usd)
+
+        2026-03-29 ogt: 加入 Token/Cost 追蹤
         """
         # 生成快取鍵 (基於 prompt + alert_context hash)
         context_hash = ""
@@ -711,13 +739,15 @@ class OpenClawService:
                     f"{cached_data['provider']}_cached",
                     True,
                     True,  # from_cache
+                    0,     # tokens (cache hit, no new tokens)
+                    0.0,   # cost (cache hit, no cost)
                 )
         except Exception as e:
             logger.warning("llm_cache_read_failed", error=str(e))
 
         # 2. Cache Miss - 呼叫 LLM
         logger.info("llm_cache_miss", cache_key=cache_key[:20])
-        response, provider, success = await self._call_with_fallback(
+        response, provider, success, total_tokens, cost_usd = await self._call_with_fallback(
             prompt, alert_context, signoz_metrics
         )
 
@@ -744,7 +774,7 @@ class OpenClawService:
             except Exception as e:
                 logger.warning("llm_cache_write_failed", error=str(e))
 
-        return response, provider, success, False  # from_cache=False
+        return response, provider, success, False, total_tokens, cost_usd  # from_cache=False
 
     # =========================================================================
     # Public LLM Interface (ILLMProvider Protocol)
@@ -773,19 +803,23 @@ class OpenClawService:
         prompt: str,
         alert_context: dict | None = None,
         signoz_metrics: GoldMetrics | None = None,
-    ) -> tuple[str, str, bool]:
+    ) -> tuple[str, str, bool, int, float]:
         """
         依 AI_FALLBACK_ORDER 順序呼叫 AI
 
         若 MOCK_MODE=True，直接回傳模擬結果。
         若所有 Provider 失敗，fallback 到 Mock。
 
+        Returns:
+            tuple: (response, provider, success, total_tokens, cost_usd)
+
         Phase 15.1: 整合 Langfuse LLMOps 追蹤
+        2026-03-29 ogt: 加入 Token/Cost 追蹤
         """
         # Mock Mode: 開發測試用
         if settings.MOCK_MODE:
             logger.info("mock_mode_enabled", using="mock_llm")
-            return self._generate_mock_response(alert_context or {}, signoz_metrics), "mock", True
+            return self._generate_mock_response(alert_context or {}, signoz_metrics), "mock", True, 0, 0.0
 
         # Phase 15.1 + 15.3: Langfuse 追蹤整合 + SignOz Deep Linking
         with langfuse_trace(
@@ -833,10 +867,14 @@ class OpenClawService:
                 start_time = time.time()
                 model_name = self._get_model_name(provider)
 
+                # 2026-03-29 ogt: Gemini 回傳 4 值 (含 token/cost)，其他 Provider 補 0
+                total_tokens = 0
+                cost_usd = 0.0
+
                 if provider == "ollama":
                     response, success = await self._call_ollama(prompt)
                 elif provider == "gemini":
-                    response, success = await self._call_gemini(prompt)
+                    response, success, total_tokens, cost_usd = await self._call_gemini(prompt)
                 elif provider == "claude":
                     response, success = await self._call_claude(prompt)
                 else:
@@ -855,21 +893,29 @@ class OpenClawService:
                         "success": success,
                         "latency_ms": round(latency_ms, 2),
                         "provider": provider,
+                        "total_tokens": total_tokens,
+                        "cost_usd": cost_usd,
                     },
                 )
 
                 if success:
-                    logger.info("ai_provider_success", provider=provider, latency_ms=latency_ms)
+                    logger.info(
+                        "ai_provider_success",
+                        provider=provider,
+                        latency_ms=latency_ms,
+                        total_tokens=total_tokens,
+                        cost_usd=f"${cost_usd:.6f}",
+                    )
                     # Langfuse: 記錄成功評分
                     trace.score(name="provider_success", value=1.0, comment=f"Success via {provider}")
-                    return response, provider, True
+                    return response, provider, True, total_tokens, cost_usd
 
                 logger.warning("ai_provider_failed_fallback", provider=provider, latency_ms=latency_ms)
 
             # 所有 Provider 失敗時，fallback 到 Mock (優雅降級)
             logger.warning("all_providers_failed_using_mock", fallback="mock_llm")
             trace.score(name="provider_success", value=0.0, comment="All providers failed, using mock")
-            return self._generate_mock_response(alert_context or {}, signoz_metrics), "mock_fallback", True
+            return self._generate_mock_response(alert_context or {}, signoz_metrics), "mock_fallback", True, 0, 0.0
 
     def _get_model_name(self, provider: str) -> str:
         """取得 provider 對應的模型名稱 (從 ModelRegistry)"""
@@ -977,7 +1023,7 @@ class OpenClawService:
     async def analyze_alert(
         self,
         alert_context: dict,
-    ) -> tuple[LLMAnalysisResult | None, str, str, GoldMetrics | None, str]:
+    ) -> tuple[LLMAnalysisResult | None, str, str, GoldMetrics | None, str, int, float]:
         """
         分析告警並產生 RCA 結果 (含 SignOz 整合)
 
@@ -985,7 +1031,9 @@ class OpenClawService:
             alert_context: 告警上下文 (alert_type, severity, target_resource, etc.)
 
         Returns:
-            (analysis_result, ai_provider, raw_response, signoz_metrics, signoz_trace_url)
+            (analysis_result, ai_provider, raw_response, signoz_metrics, signoz_trace_url, total_tokens, cost_usd)
+
+        2026-03-29 ogt: 加入 Token/Cost 追蹤
         """
         # Step 0: 擷取 SignOz 上下文
         service_name = alert_context.get("target_resource", "unknown")
@@ -1018,7 +1066,7 @@ Trace URL: {signoz_trace_url}
         )
 
         # 呼叫 LLM (使用快取層保護算力)
-        raw_response, provider, success, from_cache = await self._call_with_cache(
+        raw_response, provider, success, from_cache, total_tokens, cost_usd = await self._call_with_cache(
             full_prompt,
             alert_context,
             signoz_metrics,
@@ -1027,7 +1075,7 @@ Trace URL: {signoz_trace_url}
 
         if not success:
             logger.error("openclaw_all_providers_failed")
-            return None, provider, raw_response, signoz_metrics, signoz_trace_url
+            return None, provider, raw_response, signoz_metrics, signoz_trace_url, 0, 0.0
 
         if from_cache:
             logger.info("openclaw_using_cached_response", provider=provider)
@@ -1056,7 +1104,7 @@ Trace URL: {signoz_trace_url}
                 raw_response=raw_response[:300],
             )
 
-        return result, provider, raw_response, signoz_metrics, signoz_trace_url
+        return result, provider, raw_response, signoz_metrics, signoz_trace_url, total_tokens, cost_usd
 
     # =========================================================================
     # Phase 6.4: LLM Proposal Generation
diff --git a/apps/api/src/services/telegram_gateway.py b/apps/api/src/services/telegram_gateway.py
index 67ed0c4e2..9a961b941 100644
--- a/apps/api/src/services/telegram_gateway.py
+++ b/apps/api/src/services/telegram_gateway.py
@@ -139,6 +139,9 @@ class TelegramMessage:
     signoz_metrics: SignOzMetricsBlock | None = None
     signoz_trace_url: str = ""  # 動態時間參數 URL
     auto_tuning_command: str = ""  # kubectl 調優指令
+    # 2026-03-29 ogt: AI Token/Cost 追蹤
+    ai_tokens: int = 0          # LLM Token 使用量
+    ai_cost: float = 0.0        # LLM 成本 (USD)
 
     def format(self) -> str:
         """
@@ -190,6 +193,11 @@ class TelegramMessage:
         safe_action = html.escape(self.suggested_action[:35])
         safe_downtime = html.escape(self.estimated_downtime)
 
+        # 2026-03-29 ogt: AI Token/Cost 顯示
+        ai_cost_display = ""
+        if self.ai_tokens > 0 or self.ai_cost > 0:
+            ai_cost_display = f"💰 Tokens: {self.ai_tokens:,} / ${self.ai_cost:.4f}\n"
+
         # 組裝訊息
         message = (
             f"═══════════════════════════\n"
@@ -201,6 +209,7 @@ class TelegramMessage:
             f"🤖 <b>AI 仲裁判定</b>\n"
             f"👥 責任: {resp_display}\n"
             f"📊 信心: {conf_emoji} {confidence_pct}%\n"
+            f"{ai_cost_display}"
             f"💡 原因: {safe_root_cause}\n"
             f"{signoz_block}"
             f"━━━━━━━━━━━━━━━━━━━\n"
@@ -425,6 +434,9 @@ class TelegramGateway:
         signoz_latency_trend: str = "stable",
         signoz_trace_url: str = "",
         auto_tuning_command: str = "",
+        # 2026-03-29 ogt: AI Token/Cost 追蹤
+        ai_tokens: int = 0,
+        ai_cost: float = 0.0,
     ) -> dict:
         """
         推送待簽核卡片到 Telegram (v7.0 含 SignOz 整合)
@@ -463,7 +475,7 @@ class TelegramGateway:
                 trace_url=signoz_trace_url,
             )
 
-        # 建立訊息結構 (含 AI 仲裁 + SignOz)
+        # 建立訊息結構 (含 AI 仲裁 + SignOz + Token/Cost)
         message = TelegramMessage(
             status_emoji=emoji,
             risk_level=risk_level.upper(),
@@ -478,6 +490,9 @@ class TelegramGateway:
             signoz_metrics=signoz_metrics,
             signoz_trace_url=signoz_trace_url,
             auto_tuning_command=auto_tuning_command,
+            # 2026-03-29 ogt: AI Token/Cost 追蹤
+            ai_tokens=ai_tokens,
+            ai_cost=ai_cost,
         )
 
         # 格式化訊息
diff --git a/k8s/awoooi-prod/04-configmap.yaml b/k8s/awoooi-prod/04-configmap.yaml
index 39edac66a..d828d061a 100644
--- a/k8s/awoooi-prod/04-configmap.yaml
+++ b/k8s/awoooi-prod/04-configmap.yaml
@@ -34,9 +34,10 @@ data:
   CORS_ORIGINS: '["https://awoooi.wooo.work","http://localhost:3000","http://localhost:3001"]'
 
   # AI 配置 (JSON array 格式 for pydantic-settings)
-  # 2026-03-28: 已切回 Ollama 優先 (成本最佳化)
-  # 備援順序: Ollama($0) → Gemini(~$0.001) → Claude(~$0.008)
-  AI_FALLBACK_ORDER: '["ollama","gemini","claude"]'
+  # 2026-03-29 ogt: 切換 Gemini 優先 (確認 AI 仲裁運作正常)
+  # 備援順序: Gemini(~$0.001) → Ollama($0) → Claude(~$0.008)
+  # Rate Limiter: RPM=10, 每日 500 次, Token 100K
+  AI_FALLBACK_ORDER: '["gemini","ollama","claude"]'
   AI_CACHE_TTL: "3600"
 
   # 快取 TTL (秒)