diff --git a/src-tauri/src/services/balance.rs b/src-tauri/src/services/balance.rs
index dce56fed8..edeba3888 100644
--- a/src-tauri/src/services/balance.rs
+++ b/src-tauri/src/services/balance.rs
@@ -72,7 +72,7 @@ async fn query_deepseek(api_key: &str) -> UsageResult {
         .get("https://api.deepseek.com/user/balance")
         .header("Authorization", format!("Bearer {api_key}"))
         .header("Accept", "application/json")
-        .timeout(Duration::from_secs(10))
+        .timeout(Duration::from_secs(15))
         .send()
         .await;
 
@@ -144,7 +144,7 @@ async fn query_stepfun(api_key: &str) -> UsageResult {
         .get("https://api.stepfun.com/v1/accounts")
         .header("Authorization", format!("Bearer {api_key}"))
         .header("Accept", "application/json")
-        .timeout(Duration::from_secs(10))
+        .timeout(Duration::from_secs(15))
         .send()
         .await;
 
@@ -203,7 +203,7 @@ async fn query_siliconflow(api_key: &str, is_cn: bool) -> UsageResult {
         .get(&url)
         .header("Authorization", format!("Bearer {api_key}"))
         .header("Accept", "application/json")
-        .timeout(Duration::from_secs(10))
+        .timeout(Duration::from_secs(15))
         .send()
         .await;
 
@@ -267,7 +267,7 @@ async fn query_openrouter(api_key: &str) -> UsageResult {
         .get("https://openrouter.ai/api/v1/credits")
         .header("Authorization", format!("Bearer {api_key}"))
         .header("Accept", "application/json")
-        .timeout(Duration::from_secs(10))
+        .timeout(Duration::from_secs(15))
         .send()
         .await;
 
@@ -327,7 +327,7 @@ async fn query_novita(api_key: &str) -> UsageResult {
         .get("https://api.novita.ai/v3/user/balance")
         .header("Authorization", format!("Bearer {api_key}"))
         .header("Accept", "application/json")
-        .timeout(Duration::from_secs(10))
+        .timeout(Duration::from_secs(15))
         .send()
         .await;
 
diff --git a/src-tauri/src/services/coding_plan.rs b/src-tauri/src/services/coding_plan.rs
index bb547f207..25c2f009b 100644
--- a/src-tauri/src/services/coding_plan.rs
+++ b/src-tauri/src/services/coding_plan.rs
@@ -95,7 +95,7 @@ async fn query_kimi(api_key: &str) -> SubscriptionQuota {
         .get("https://api.kimi.com/coding/v1/usages")
         .header("Authorization", format!("Bearer {api_key}"))
         .header("Accept", "application/json")
-        .timeout(std::time::Duration::from_secs(10))
+        .timeout(std::time::Duration::from_secs(15))
         .send()
         .await;
 
@@ -308,7 +308,7 @@ async fn query_zhipu(base_url: &str, api_key: &str) -> SubscriptionQuota {
         .header("Authorization", api_key) // 注意：智谱不加 Bearer 前缀
         .header("Content-Type", "application/json")
         .header("Accept-Language", "en-US,en")
-        .timeout(std::time::Duration::from_secs(10))
+        .timeout(std::time::Duration::from_secs(15))
         .send()
         .await;
 
@@ -391,7 +391,7 @@ async fn query_minimax(api_key: &str, is_cn: bool) -> SubscriptionQuota {
         .get(&url)
         .header("Authorization", format!("Bearer {api_key}"))
         .header("Content-Type", "application/json")
-        .timeout(std::time::Duration::from_secs(10))
+        .timeout(std::time::Duration::from_secs(15))
         .send()
         .await;
 
@@ -463,7 +463,7 @@ async fn query_zenmux(base_url: &str, api_key: &str) -> SubscriptionQuota {
         .get(base_url)
         .header("Authorization", format!("Bearer {api_key}"))
         .header("Accept", "application/json")
-        .timeout(std::time::Duration::from_secs(10))
+        .timeout(std::time::Duration::from_secs(15))
         .send()
         .await;
 
@@ -666,7 +666,8 @@ pub async fn get_coding_plan_quota(
             success: false,
             tiers: vec![],
             extra_usage: None,
-            error: None,
+            // 与 balance::get_balance 一致：给出明确错误，避免 footer 显示无信息的失败
+            error: Some("API key is empty".to_string()),
             queried_at: None,
         });
     }
@@ -681,9 +682,10 @@ pub async fn get_coding_plan_quota(
                 success: false,
                 tiers: vec![],
                 extra_usage: None,
-                error: None,
+                // 域名未命中已知套餐供应商（如第三方中转站）：给出明确错误而非静默失败
+                error: Some("Unknown coding plan provider".to_string()),
                 queried_at: None,
-            })
+            });
         }
     };
 
diff --git a/src-tauri/src/services/subscription.rs b/src-tauri/src/services/subscription.rs
index c7a6fa66c..092100349 100644
--- a/src-tauri/src/services/subscription.rs
+++ b/src-tauri/src/services/subscription.rs
@@ -328,7 +328,7 @@ async fn query_claude_quota(access_token: &str) -> SubscriptionQuota {
         .header("Authorization", format!("Bearer {access_token}"))
         .header("anthropic-beta", "oauth-2025-04-20")
         .header("Accept", "application/json")
-        .timeout(std::time::Duration::from_secs(10))
+        .timeout(std::time::Duration::from_secs(15))
         .send()
         .await;
 
@@ -668,7 +668,7 @@ pub(crate) async fn query_codex_quota(
         req = req.header("ChatGPT-Account-Id", id);
     }
 
-    let resp = match req.timeout(std::time::Duration::from_secs(10)).send().await {
+    let resp = match req.timeout(std::time::Duration::from_secs(15)).send().await {
         Ok(r) => r,
         Err(e) => {
             return SubscriptionQuota::error(
@@ -964,7 +964,7 @@ async fn refresh_gemini_token(refresh_token: &str) -> Option<String> {
             ("refresh_token", refresh_token),
             ("grant_type", "refresh_token"),
         ])
-        .timeout(std::time::Duration::from_secs(10))
+        .timeout(std::time::Duration::from_secs(15))
         .send()
         .await
         .ok()?;
@@ -1048,7 +1048,7 @@ async fn query_gemini_quota(access_token: &str) -> SubscriptionQuota {
                 "pluginType": "GEMINI"
             }
         }))
-        .timeout(std::time::Duration::from_secs(10))
+        .timeout(std::time::Duration::from_secs(15))
         .send()
         .await;
 
@@ -1109,7 +1109,7 @@ async fn query_gemini_quota(access_token: &str) -> SubscriptionQuota {
         .header("Authorization", format!("Bearer {access_token}"))
         .header("Content-Type", "application/json")
         .json(&quota_body)
-        .timeout(std::time::Duration::from_secs(10))
+        .timeout(std::time::Duration::from_secs(15))
         .send()
         .await;
 
diff --git a/src/lib/query/queries.ts b/src/lib/query/queries.ts
index ef0711c36..28cba7556 100644
--- a/src/lib/query/queries.ts
+++ b/src/lib/query/queries.ts
@@ -1,3 +1,4 @@
+import { useRef } from "react";
 import {
   useQuery,
   type UseQueryResult,
@@ -99,6 +100,112 @@ export interface UseUsageQueryOptions {
   autoQueryInterval?: number; // 自动查询间隔（分钟），0 表示禁用
 }
 
+/** 最近一次成功的用量结果快照（keep-last-good 用）。 */
+export interface LastGoodUsage {
+  data: UsageResult;
+  at: number; // 该成功结果的获取时刻（ms）
+}
+
+/** 在最近一次成功后多久内，失败仍继续展示该成功值。 */
+export const KEEP_LAST_GOOD_MS = 10 * 60 * 1000; // 10 分钟
+
+/**
+ * 判断一次用量查询失败是否属于"瞬时/网络类"（可被 keep-last-good 短暂掩盖）。
+ *
+ * 仅瞬时失败才允许继续展示上一次成功；**确定性失败**（鉴权失败、空 API Key、
+ * 未知供应商、4xx、脚本/解析错误等）必须立即透出——用户改/删凭据后要马上看到，
+ * 否则会一直显示过期额度直到窗口结束。
+ *
+ * 采用**白名单**：只认后端稳定的网络类错误前缀 + HTTP 5xx，失败安全——任何未识别
+ * 的错误一律按"非瞬时"立即透出，绝不误掩盖确定性失败。需与后端错误文案保持同步：
+ * - 原生 balance/coding_plan/subscription：reqwest send 失败/超时 → `"Network error: …"`；
+ *   上游非 2xx → `"API error (HTTP <code>…)"`
+ * - JS 脚本 usage_script：`request_failed` → "请求失败/Request failed"、
+ *   `read_response_failed` → "读取响应失败/Failed to read response"（仅 zh/en 两种本地化）；
+ *   上游非 2xx → `"HTTP <code> …"`
+ *
+ * HTTP 状态：**5xx**（服务端错误，通常瞬时）归为 transient；**4xx**（鉴权/客户端
+ * 错误，如 401/403/404/429）保持确定性，立即透出。
+ */
+export function isTransientUsageError(result: UsageResult): boolean {
+  if (result.success) return false;
+  const e = result.error?.toLowerCase() ?? "";
+  if (!e) return false;
+
+  // 网络类（send 失败/超时/读取响应失败）
+  if (
+    e.includes("network error") || // 原生路径
+    e.includes("request failed") || // JS 脚本 (en)
+    e.includes("请求失败") || // JS 脚本 (zh)
+    e.includes("failed to read response") || // JS 脚本 (en)
+    e.includes("读取响应失败") // JS 脚本 (zh)
+  ) {
+    return true;
+  }
+
+  // HTTP 状态码：5xx 视为瞬时，4xx 视为确定性。错误文案里第一处 "HTTP <code>" 即为
+  // 上游状态码（原生 "API error (HTTP 500…)"、JS 脚本 "HTTP 500 …"）。
+  const httpMatch = e.match(/http\s+(\d{3})/);
+  if (httpMatch) {
+    const status = Number(httpMatch[1]);
+    return status >= 500 && status <= 599;
+  }
+
+  return false;
+}
+
+/**
+ * Keep-last-good 的纯决策函数（无 ref、无时钟，`now` 注入以便测试）。
+ *
+ * 用量端点面向跨境/第三方，单次网络抖动会让后端返回 `Ok(success:false)` 当作"新数据"
+ * 覆盖掉上一次成功，使卡片瞬间翻红——这是"一会成功一会失败"观感的主因。
+ * 策略：当失败是**瞬时/网络类**（见 [`isTransientUsageError`]）且最近一次成功在
+ * `keepMs` 内时，不抹掉它，继续展示该成功值，并把 `lastQueriedAt` 指向该成功的时刻
+ * （相对时间自然走到"10 分钟前"后过期翻红）；超出窗口、或从无成功记录时照常展示失败。
+ *
+ * **确定性失败**（鉴权/空 key/未知供应商/4xx 等）不仅立即透出，还会**清空 `lastGood`**：
+ * 旧成功快照已不可信，否则随后一次网络抖动会把"配置/鉴权已失效"的旧额度重新复活。
+ *
+ * 注：会 reject 的传输层失败（Copilot/DB）react-query 本就保留上次 `data`，这里
+ * 主要修的是 `Ok(success:false)` 这条覆盖路径。
+ */
+export function resolveDisplayUsage(
+  raw: UsageResult | undefined,
+  dataUpdatedAt: number,
+  prevLastGood: LastGoodUsage | null,
+  now: number,
+  keepMs: number = KEEP_LAST_GOOD_MS,
+): {
+  data: UsageResult | undefined;
+  lastQueriedAt: number | null;
+  lastGood: LastGoodUsage | null;
+} {
+  let lastGood = prevLastGood;
+  if (raw?.success) {
+    // 成功：刷新快照
+    lastGood = { data: raw, at: dataUpdatedAt || now };
+  } else if (raw && !isTransientUsageError(raw)) {
+    // 确定性失败（鉴权/空 key/未知供应商/4xx 等）：旧成功快照已不可信，丢弃它，
+    // 避免后续一次网络抖动把"配置/鉴权已失效"的旧额度重新复活。
+    lastGood = null;
+  }
+
+  let data = raw;
+  let lastQueriedAt = dataUpdatedAt || null;
+  if (
+    raw &&
+    !raw.success &&
+    isTransientUsageError(raw) && // 仅瞬时/网络类失败才掩盖；确定性失败立即透出
+    lastGood &&
+    now - lastGood.at < keepMs
+  ) {
+    data = lastGood.data;
+    lastQueriedAt = lastGood.at;
+  }
+
+  return { data, lastQueriedAt, lastGood };
+}
+
 export const useUsageQuery = (
   providerId: string,
   appId: AppId,
@@ -123,14 +230,31 @@ export const useUsageQuery = (
         : false,
     refetchIntervalInBackground: true, // 后台也继续定时查询
     refetchOnWindowFocus: false,
-    retry: false,
+    // 用量查询面向跨境/第三方端点，单次网络抖动或瞬时 5xx 不应直接判失败。
+    // 重试一次以吸收瞬时故障（与 useSubscriptionQuota 的 retry:1 保持一致）。
+    // 注意：原生 balance/coding_plan 路径把网络错误折叠成 Ok(success:false)，
+    // 这类不会触发 react-query 重试；本项主要覆盖会 reject 的传输层失败（Copilot/DB 等）。
+    retry: 1,
+    retryDelay: 1500,
     staleTime, // 使用动态计算的缓存时间
     gcTime: 10 * 60 * 1000, // 缓存保留 10 分钟（组件卸载后）
   });
 
+  // Keep-last-good：失败时在 10 分钟窗口内继续展示上一次成功值（见 resolveDisplayUsage）。
+  // 每个 hook 实例各持一份 ref（按卡片维度）；ref 写入是幂等的（同份成功重复写无副作用）。
+  const lastGoodRef = useRef<LastGoodUsage | null>(null);
+  const { data, lastQueriedAt, lastGood } = resolveDisplayUsage(
+    query.data,
+    query.dataUpdatedAt,
+    lastGoodRef.current,
+    Date.now(),
+  );
+  lastGoodRef.current = lastGood;
+
   return {
     ...query,
-    lastQueriedAt: query.dataUpdatedAt || null,
+    data,
+    lastQueriedAt,
   };
 };
 
diff --git a/tests/lib/keepLastGoodUsage.test.ts b/tests/lib/keepLastGoodUsage.test.ts
new file mode 100644
index 000000000..e696f9f26
--- /dev/null
+++ b/tests/lib/keepLastGoodUsage.test.ts
@@ -0,0 +1,176 @@
+import { describe, it, expect } from "vitest";
+import {
+  resolveDisplayUsage,
+  isTransientUsageError,
+  KEEP_LAST_GOOD_MS,
+  type LastGoodUsage,
+} from "@/lib/query/queries";
+import type { UsageResult } from "@/types";
+
+// keep-last-good 的纯决策逻辑：仅"瞬时/网络类"失败才在 KEEP_LAST_GOOD_MS 窗口内继续
+// 展示上一次成功；确定性失败（鉴权/空 key/未知供应商等）必须立即透出。
+
+const ok = (remaining: number): UsageResult => ({
+  success: true,
+  data: [{ remaining, unit: "USD" }],
+});
+// 默认用网络类错误（瞬时），需要确定性失败时显式传入。
+const fail = (error = "Network error: connection reset"): UsageResult => ({
+  success: false,
+  error,
+});
+
+const T0 = 1_000_000_000_000; // 任意基准时刻（ms）
+
+describe("isTransientUsageError", () => {
+  it("网络类失败 → 瞬时（true）", () => {
+    expect(isTransientUsageError(fail("Network error: timed out"))).toBe(true);
+    expect(isTransientUsageError(fail("Request failed: timed out"))).toBe(true);
+    expect(isTransientUsageError(fail("请求失败: 连接超时"))).toBe(true);
+    expect(isTransientUsageError(fail("Failed to read response: eof"))).toBe(
+      true,
+    );
+    expect(isTransientUsageError(fail("读取响应失败: eof"))).toBe(true);
+  });
+
+  it("确定性失败 → 非瞬时（false），必须立即透出", () => {
+    expect(
+      isTransientUsageError(fail("Authentication failed (HTTP 401)")),
+    ).toBe(false);
+    expect(isTransientUsageError(fail("API key is empty"))).toBe(false);
+    expect(isTransientUsageError(fail("Unknown balance provider"))).toBe(false);
+    expect(isTransientUsageError(fail("Unknown coding plan provider"))).toBe(
+      false,
+    );
+    expect(isTransientUsageError(fail("API error (HTTP 400): bad"))).toBe(
+      false,
+    );
+    expect(isTransientUsageError(fail("Failed to parse response: x"))).toBe(
+      false,
+    );
+  });
+
+  it("HTTP 5xx → 瞬时（true）；4xx → 非瞬时（false）", () => {
+    expect(isTransientUsageError(fail("API error (HTTP 500): oops"))).toBe(
+      true,
+    );
+    expect(
+      isTransientUsageError(fail("HTTP 503 Service Unavailable : x")),
+    ).toBe(true);
+    expect(
+      isTransientUsageError(fail("API error (HTTP 502): bad gateway")),
+    ).toBe(true);
+    expect(
+      isTransientUsageError(fail("API error (HTTP 429): rate limited")),
+    ).toBe(false);
+    expect(
+      isTransientUsageError(fail("Authentication failed (HTTP 403)")),
+    ).toBe(false);
+  });
+
+  it("成功 / 无错误信息 → false", () => {
+    expect(isTransientUsageError(ok(1))).toBe(false);
+    expect(isTransientUsageError({ success: false })).toBe(false);
+  });
+});
+
+describe("resolveDisplayUsage (keep-last-good)", () => {
+  it("成功结果：原样展示并记录为 lastGood，lastQueriedAt=获取时刻", () => {
+    const success = ok(42);
+    const r = resolveDisplayUsage(success, T0, null, T0);
+    expect(r.data).toBe(success);
+    expect(r.lastQueriedAt).toBe(T0);
+    expect(r.lastGood).toEqual({ data: success, at: T0 });
+  });
+
+  it("瞬时失败 + 窗口内有上次成功：继续展示成功值，lastQueriedAt 指向成功时刻", () => {
+    const prev: LastGoodUsage = { data: ok(42), at: T0 };
+    const now = T0 + KEEP_LAST_GOOD_MS - 1; // 刚好仍在窗口内
+    const r = resolveDisplayUsage(fail(), now, prev, now);
+    expect(r.data).toBe(prev.data); // 展示的是上次成功
+    expect(r.lastQueriedAt).toBe(T0); // 时间戳反映成功的年龄
+    expect(r.lastGood).toBe(prev); // 失败不更新 lastGood
+  });
+
+  it("瞬时失败 + 上次成功已过期（>= 窗口）：展示失败本身", () => {
+    const prev: LastGoodUsage = { data: ok(42), at: T0 };
+    const now = T0 + KEEP_LAST_GOOD_MS; // 边界：恰好到 10 分钟即过期
+    const failure = fail();
+    const r = resolveDisplayUsage(failure, now, prev, now);
+    expect(r.data).toBe(failure);
+    expect(r.lastQueriedAt).toBe(now);
+    expect(r.lastGood).toBe(prev);
+  });
+
+  it("确定性失败（鉴权/空 key/未知供应商）：即使窗口内有上次成功也立即透出，并清空 lastGood", () => {
+    const prev: LastGoodUsage = { data: ok(42), at: T0 };
+    const now = T0 + 1000; // 远在窗口内
+    for (const failure of [
+      fail("Authentication failed (HTTP 401)"),
+      fail("API key is empty"),
+      fail("Unknown coding plan provider"),
+    ]) {
+      const r = resolveDisplayUsage(failure, now, prev, now);
+      expect(r.data).toBe(failure); // 不掩盖 → 透出确定性失败
+      expect(r.lastQueriedAt).toBe(now);
+      expect(r.lastGood).toBeNull(); // 旧快照已不可信 → 清空，防止后续被复活
+    }
+  });
+
+  it("确定性失败清空 lastGood：随后的网络抖动不会复活旧成功", () => {
+    // 成功 → 记录 lastGood
+    const afterSuccess = resolveDisplayUsage(ok(42), T0, null, T0);
+    expect(afterSuccess.lastGood).not.toBeNull();
+    // 401 鉴权失败 → 透出失败并清空 lastGood
+    const afterAuthFail = resolveDisplayUsage(
+      fail("Authentication failed (HTTP 401)"),
+      T0 + 1000,
+      afterSuccess.lastGood,
+      T0 + 1000,
+    );
+    expect(afterAuthFail.lastGood).toBeNull();
+    // 随后一次网络抖动（瞬时）→ lastGood 已空 → 不复活旧成功，照常透出失败
+    const netFail = fail();
+    const afterBlip = resolveDisplayUsage(
+      netFail,
+      T0 + 2000,
+      afterAuthFail.lastGood,
+      T0 + 2000,
+    );
+    expect(afterBlip.data).toBe(netFail);
+    expect(afterBlip.lastGood).toBeNull();
+  });
+
+  it("瞬时失败 + 从无成功记录：展示失败本身", () => {
+    const failure = fail();
+    const now = T0 + 5000;
+    const r = resolveDisplayUsage(failure, now, null, now);
+    expect(r.data).toBe(failure);
+    expect(r.lastQueriedAt).toBe(now);
+    expect(r.lastGood).toBeNull();
+  });
+
+  it("新的成功覆盖旧的 lastGood", () => {
+    const prev: LastGoodUsage = { data: ok(42), at: T0 };
+    const fresh = ok(7);
+    const now = T0 + 60_000;
+    const r = resolveDisplayUsage(fresh, now, prev, now);
+    expect(r.data).toBe(fresh);
+    expect(r.lastGood).toEqual({ data: fresh, at: now });
+  });
+
+  it("加载中（raw=undefined）：data 为 undefined，lastGood 不变", () => {
+    const prev: LastGoodUsage = { data: ok(42), at: T0 };
+    const r = resolveDisplayUsage(undefined, 0, prev, T0 + 1000);
+    expect(r.data).toBeUndefined();
+    expect(r.lastQueriedAt).toBeNull();
+    expect(r.lastGood).toBe(prev);
+  });
+
+  it("dataUpdatedAt 为 0 的成功：用注入的 now 作为获取时刻", () => {
+    const success = ok(1);
+    const now = T0 + 123;
+    const r = resolveDisplayUsage(success, 0, null, now);
+    expect(r.lastGood).toEqual({ data: success, at: now });
+  });
+});