diff --git a/backend/internal/service/antigravity_gateway_service.go b/backend/internal/service/antigravity_gateway_service.go
index 7c4193a3..30675b02 100644
--- a/backend/internal/service/antigravity_gateway_service.go
+++ b/backend/internal/service/antigravity_gateway_service.go
@@ -44,9 +44,10 @@ const (
 
 	// MODEL_CAPACITY_EXHAUSTED 专用重试参数
 	// 模型容量不足时，所有账号共享同一容量池，切换账号无意义
-	// 使用固定 1s 间隔重试，最多重试 60 次
-	antigravityModelCapacityRetryMaxAttempts = 60
+	// 使用指数退避策略重试，最多重试 10 次（而非 60 次）
+	antigravityModelCapacityRetryMaxAttempts = 10
 	antigravityModelCapacityRetryWait        = 1 * time.Second
+	antigravityModelCapacityRetryMaxWait     = 32 * time.Second // 指数退避上限
 
 	// Google RPC 状态和类型常量
 	googleRPCStatusResourceExhausted      = "RESOURCE_EXHAUSTED"
@@ -307,7 +308,7 @@ func (s *AntigravityGatewayService) handleSmartRetry(p antigravityRetryLoopParam
 		var lastRetryResp *http.Response
 		var lastRetryBody []byte
 
-		// MODEL_CAPACITY_EXHAUSTED 使用独立的重试参数（60 次，固定 1s 间隔）
+		// MODEL_CAPACITY_EXHAUSTED 使用独立的重试参数（10 次，指数退避）
 		maxAttempts := antigravitySmartRetryMaxAttempts
 		if isModelCapacityExhausted {
 			maxAttempts = antigravityModelCapacityRetryMaxAttempts
@@ -334,10 +335,29 @@ func (s *AntigravityGatewayService) handleSmartRetry(p antigravityRetryLoopParam
 		}
 
 		for attempt := 1; attempt <= maxAttempts; attempt++ {
-			log.Printf("%s status=%d oauth_smart_retry attempt=%d/%d delay=%v model=%s account=%d",
-				p.prefix, resp.StatusCode, attempt, maxAttempts, waitDuration, modelName, p.account.ID)
+			// 计算本次重试的等待时间
+			var currentWaitDuration time.Duration
+			if isModelCapacityExhausted {
+				// 使用指数退避：1s, 2s, 4s, 8s, 16s, 32s, ...
+				currentWaitDuration = waitDuration * time.Duration(1<<(attempt-1))
+				if currentWaitDuration > antigravityModelCapacityRetryMaxWait {
+					currentWaitDuration = antigravityModelCapacityRetryMaxWait
+				}
+				// 添加随机抖动（±10%）避免羊群效应
+				jitter := time.Duration(mathrand.Int63n(int64(currentWaitDuration / 5)))
+				if mathrand.Intn(2) == 0 {
+					currentWaitDuration += jitter
+				} else {
+					currentWaitDuration -= jitter
+				}
+			} else {
+				currentWaitDuration = waitDuration
+			}
 
-			timer := time.NewTimer(waitDuration)
+			log.Printf("%s status=%d oauth_smart_retry attempt=%d/%d delay=%v model=%s account=%d",
+				p.prefix, resp.StatusCode, attempt, maxAttempts, currentWaitDuration, modelName, p.account.ID)
+
+			timer := time.NewTimer(currentWaitDuration)
 			select {
 			case <-p.ctx.Done():
 				timer.Stop()