diff --git a/backend/internal/handler/admin/account_ollama_cloud_usage_test.go b/backend/internal/handler/admin/account_ollama_cloud_usage_test.go
index 89f2aa44ab..38db30a2c1 100644
--- a/backend/internal/handler/admin/account_ollama_cloud_usage_test.go
+++ b/backend/internal/handler/admin/account_ollama_cloud_usage_test.go
@@ -273,4 +273,5 @@ func TestGetOllamaCloudUsageSettingsHandlerSuccess(t *testing.T) {
 	require.Equal(t, http.StatusOK, recorder.Code)
 	require.Contains(t, recorder.Body.String(), `"enabled":false`)
 	require.Contains(t, recorder.Body.String(), `"interval_minutes":60`)
+	require.Contains(t, recorder.Body.String(), `"debounce_minutes":1`)
 }
diff --git a/backend/internal/repository/account_repo_ollama_cloud_usage.go b/backend/internal/repository/account_repo_ollama_cloud_usage.go
index 78d14d2097..9d06c44ef1 100644
--- a/backend/internal/repository/account_repo_ollama_cloud_usage.go
+++ b/backend/internal/repository/account_repo_ollama_cloud_usage.go
@@ -354,9 +354,31 @@ func canonicalJSON(raw string) string {
 	return string(encoded)
 }
 
-// ListDueOllamaCloudUsageAccounts returns at most one due representative per
-// exact API key before hydration, preventing one shared group from consuming a
-// whole runner cycle.
+// ollamaCloudUsageParseRFC3339SQL reuses the verified RFC3339(/Nano) parse path
+// for a snapshot timestamp expression. Invalid or missing values fail open to NULL.
+func ollamaCloudUsageParseRFC3339SQL(expression string) string {
+	return `CASE
+		WHEN ` + expression + ` IS NULL THEN NULL
+		WHEN ` + expression + ` ~ '^[0-9]{4}-[0-9]{2}-[0-9]{2}T[0-9]{2}:[0-9]{2}:[0-9]{2}(\.[0-9]+)?(Z|[+-][0-9]{2}:[0-9]{2})$'
+			THEN jsonb_path_query_first_tz(
+				to_jsonb(regexp_replace(
+					` + expression + `,
+					'(\.[0-9]{6})[0-9]+(Z|[+-][0-9]{2}:[0-9]{2})$',
+					'\1\2'
+				)),
+				'$.datetime()', '{}'::jsonb, true
+			) #>> '{}'
+		ELSE NULL
+	END`
+}
+
+// ListDueOllamaCloudUsageAccounts returns at most one activity-driven candidate
+// per exact API key. It stamps Account.LastUsedAt with the group MAX(last_used_at)
+// so the service due pure function can decide without hydrating the whole table.
+//
+// SQL only prefilters potential candidates (missing/invalid snapshot, or group
+// activity after the last successful fetch / failed attempt). Final due timing
+// (debounce, max-wait, failure backoff) is evaluated in service code.
 func (r *accountRepository) ListDueOllamaCloudUsageAccounts(ctx context.Context, now time.Time, limit int) ([]service.Account, error) {
 	if limit <= 0 {
 		return []service.Account{}, nil
@@ -364,64 +386,90 @@ func (r *accountRepository) ListDueOllamaCloudUsageAccounts(ctx context.Context,
 	if r == nil || r.sql == nil {
 		return nil, errors.New("account repository SQL executor not configured")
 	}
+	// now is retained in the signature for callers; activity due is evaluated in service.
+	_ = now
 	rows, err := r.sql.QueryContext(ctx, `
-		WITH candidates AS (
-			SELECT id, credentials ->> 'api_key' AS api_key,
-				extra #>> '{ollama_cloud_usage_snapshot,next_refresh_at}' AS next_refresh_at
+		WITH eligible AS (
+			SELECT id,
+				credentials ->> 'api_key' AS api_key,
+				last_used_at,
+				extra -> 'ollama_cloud_usage_snapshot' AS snapshot
 			FROM accounts
 			WHERE deleted_at IS NULL
 				AND status = 'active'
 				AND `+ollamaCloudUsageEligibleSQL+`
 				AND jsonb_typeof(extra -> 'ollama_cloud_usage_session') = 'string'
 				AND extra @> '{"ollama_cloud_usage_auto_refresh": true}'::jsonb
+		), group_activity AS (
+			SELECT credentials ->> 'api_key' AS api_key,
+				MAX(last_used_at) AS group_last_used_at
+			FROM accounts
+			WHERE deleted_at IS NULL
+				AND `+ollamaCloudUsageEligibleSQL+`
+				AND jsonb_typeof(credentials -> 'api_key') = 'string'
+			GROUP BY credentials ->> 'api_key'
+		), joined AS (
+			SELECT e.id, e.api_key, e.snapshot, g.group_last_used_at,
+				e.snapshot #>> '{status}' AS status,
+				e.snapshot #>> '{fetched_at}' AS fetched_at,
+				e.snapshot #>> '{last_attempt_at}' AS last_attempt_at
+			FROM eligible e
+			JOIN group_activity g ON g.api_key = e.api_key
 		), parsed AS MATERIALIZED (
-			SELECT id, api_key, next_refresh_at,
-				next_refresh_at ~ '^[0-9]{4}-[0-9]{2}-[0-9]{2}T[0-9]{2}:[0-9]{2}:[0-9]{2}(\.[0-9]+)?(Z|[+-][0-9]{2}:[0-9]{2})$' AS rfc3339_shape,
-				jsonb_path_query_first_tz(
-					to_jsonb(regexp_replace(
-						next_refresh_at,
-						'(\.[0-9]{6})[0-9]+(Z|[+-][0-9]{2}:[0-9]{2})$',
-						'\1\2'
-					)),
-					'$.datetime()', '{}'::jsonb, true
-				) #>> '{}' AS parsed_next_refresh_at
-			FROM candidates
-		), due AS (
+			SELECT id, api_key, snapshot, group_last_used_at, status,
+				`+ollamaCloudUsageParseRFC3339SQL("fetched_at")+` AS parsed_fetched_at,
+				`+ollamaCloudUsageParseRFC3339SQL("last_attempt_at")+` AS parsed_last_attempt_at
+			FROM joined
+		), candidates AS (
 			SELECT *,
-				CASE WHEN next_refresh_at IS NULL OR NOT rfc3339_shape OR parsed_next_refresh_at IS NULL THEN 0 ELSE 1 END AS due_class
+				CASE
+					WHEN snapshot IS NULL OR snapshot = 'null'::jsonb OR status IS NULL
+						OR status NOT IN ('ok', 'failed', 'unauthorized') THEN 0
+					WHEN status = 'ok' AND parsed_fetched_at IS NULL THEN 0
+					WHEN status IN ('failed', 'unauthorized') AND parsed_last_attempt_at IS NULL THEN 0
+					WHEN status = 'ok'
+						AND group_last_used_at IS NOT NULL
+						AND group_last_used_at > parsed_fetched_at::timestamptz THEN 1
+					WHEN status IN ('failed', 'unauthorized')
+						AND group_last_used_at IS NOT NULL
+						AND group_last_used_at > parsed_last_attempt_at::timestamptz THEN 1
+					ELSE NULL
+				END AS due_class
 			FROM parsed
-			WHERE next_refresh_at IS NULL
-				OR NOT rfc3339_shape
-				OR parsed_next_refresh_at IS NULL
-				OR parsed_next_refresh_at::timestamptz <= $1
 		), ranked AS (
-			SELECT *, row_number() OVER (
-				PARTITION BY api_key
-				ORDER BY due_class,
-					CASE WHEN rfc3339_shape AND parsed_next_refresh_at IS NOT NULL THEN parsed_next_refresh_at::timestamptz END NULLS FIRST,
-					id
-			) AS group_rank
-			FROM due
+			SELECT id, api_key, group_last_used_at, due_class,
+				row_number() OVER (
+					PARTITION BY api_key
+					ORDER BY due_class,
+						group_last_used_at NULLS FIRST,
+						id
+				) AS group_rank
+			FROM candidates
+			WHERE due_class IS NOT NULL
 		)
-		SELECT id
+		SELECT id, group_last_used_at
 		FROM ranked
 		WHERE group_rank = 1
-		ORDER BY due_class,
-			CASE WHEN rfc3339_shape AND parsed_next_refresh_at IS NOT NULL THEN parsed_next_refresh_at::timestamptz END NULLS FIRST,
-			id
-		LIMIT $2
-	`, now.UTC(), limit)
+		ORDER BY due_class, group_last_used_at NULLS FIRST, id
+		LIMIT $1
+	`, limit)
 	if err != nil {
 		return nil, err
 	}
 	defer func() { _ = rows.Close() }()
+	type dueRow struct {
+		id            int64
+		groupLastUsed *time.Time
+	}
+	rowsOut := make([]dueRow, 0, limit)
 	ids := make([]int64, 0, limit)
 	for rows.Next() {
-		var id int64
-		if err := rows.Scan(&id); err != nil {
+		var row dueRow
+		if err := rows.Scan(&row.id, &row.groupLastUsed); err != nil {
 			return nil, err
 		}
-		ids = append(ids, id)
+		rowsOut = append(rowsOut, row)
+		ids = append(ids, row.id)
 	}
 	if err := rows.Err(); err != nil {
 		return nil, err
@@ -430,11 +478,26 @@ func (r *accountRepository) ListDueOllamaCloudUsageAccounts(ctx context.Context,
 	if err != nil {
 		return nil, err
 	}
-	result := make([]service.Account, 0, len(accounts))
+	byID := make(map[int64]*service.Account, len(accounts))
 	for _, account := range accounts {
 		if account != nil {
-			result = append(result, *account)
+			byID[account.ID] = account
 		}
 	}
+	result := make([]service.Account, 0, len(rowsOut))
+	for _, row := range rowsOut {
+		account := byID[row.id]
+		if account == nil {
+			continue
+		}
+		// Stamp group MAX(last_used_at) for service due evaluation.
+		if row.groupLastUsed != nil {
+			ts := row.groupLastUsed.UTC()
+			account.LastUsedAt = &ts
+		} else {
+			account.LastUsedAt = nil
+		}
+		result = append(result, *account)
+	}
 	return result, nil
 }
diff --git a/backend/internal/repository/account_repo_ollama_cloud_usage_integration_test.go b/backend/internal/repository/account_repo_ollama_cloud_usage_integration_test.go
index a7dc861249..49bb79127f 100644
--- a/backend/internal/repository/account_repo_ollama_cloud_usage_integration_test.go
+++ b/backend/internal/repository/account_repo_ollama_cloud_usage_integration_test.go
@@ -23,31 +23,40 @@ func TestListDueOllamaCloudUsageAccountsOrderingLimitAndProxyHydration(t *testin
 		Username: "user", Password: "pass", Status: service.StatusActive,
 	})
 
-	createAccount := func(name, baseURL string, proxyID *int64, nextRefreshAt *time.Time) *service.Account {
+	createAccount := func(name, baseURL string, proxyID *int64, snapshot map[string]any, lastUsed *time.Time) *service.Account {
 		t.Helper()
 		extra := map[string]any{
 			service.OllamaCloudUsageSessionExtraKey:     "cipher:wos-session=fixture",
 			service.OllamaCloudUsageAutoRefreshExtraKey: true,
 		}
-		if nextRefreshAt != nil {
-			extra[service.OllamaCloudUsageSnapshotExtraKey] = map[string]any{
-				"status": service.OllamaCloudUsageStatusOK, "next_refresh_at": nextRefreshAt.UTC().Format(time.RFC3339Nano),
-			}
+		if snapshot != nil {
+			extra[service.OllamaCloudUsageSnapshotExtraKey] = snapshot
 		}
 		return mustCreateAccount(t, tx.Client(), &service.Account{
 			Name: name, Platform: service.PlatformOpenAI, Type: service.AccountTypeAPIKey,
 			Credentials: map[string]any{"api_key": name, "base_url": baseURL},
-			Extra:       extra, ProxyID: proxyID,
+			Extra:       extra, ProxyID: proxyID, LastUsedAt: lastUsed,
 		})
 	}
 
-	uppercasePath := createAccount("ollama-uppercase-path", "https://ollama.com/V1", nil, nil)
-	missingSnapshot := createAccount("ollama-due-missing", "HTTPS://WWW.OLLAMA.COM:443/v1", &proxy.ID, nil)
-	oldest := now.Add(-2 * time.Hour)
-	due := createAccount("ollama-due-oldest", "https://ollama.com", nil, &oldest)
-	future := now.Add(time.Minute)
-	_ = createAccount("ollama-not-due", "https://ollama.com", nil, &future)
-	_ = createAccount("ollama-ineligible", "https://ollama.com.evil.test", nil, nil)
+	uppercasePath := createAccount("ollama-uppercase-path", "https://ollama.com/V1", nil, nil, nil)
+	missingSnapshot := createAccount("ollama-due-missing", "HTTPS://WWW.OLLAMA.COM:443/v1", &proxy.ID, nil, nil)
+	fetched := now.Add(-2 * time.Hour)
+	activity := now.Add(-5 * time.Minute)
+	due := createAccount("ollama-due-activity", "https://ollama.com", nil, map[string]any{
+		"status":          service.OllamaCloudUsageStatusOK,
+		"fetched_at":      fetched.UTC().Format(time.RFC3339Nano),
+		"last_attempt_at": fetched.UTC().Format(time.RFC3339Nano),
+		"next_refresh_at": fetched.Add(time.Hour).UTC().Format(time.RFC3339Nano),
+	}, &activity)
+	// Success snapshot without newer activity must not be listed.
+	_ = createAccount("ollama-not-due-idle", "https://ollama.com", nil, map[string]any{
+		"status":          service.OllamaCloudUsageStatusOK,
+		"fetched_at":      now.Add(-time.Hour).UTC().Format(time.RFC3339Nano),
+		"last_attempt_at": now.Add(-time.Hour).UTC().Format(time.RFC3339Nano),
+		"next_refresh_at": now.Add(-time.Minute).UTC().Format(time.RFC3339Nano),
+	}, nil)
+	_ = createAccount("ollama-ineligible", "https://ollama.com.evil.test", nil, nil, nil)
 
 	accounts, err := repo.ListDueOllamaCloudUsageAccounts(ctx, now, 2)
 
@@ -55,49 +64,83 @@ func TestListDueOllamaCloudUsageAccountsOrderingLimitAndProxyHydration(t *testin
 	require.Len(t, accounts, 2)
 	require.Equal(t, missingSnapshot.ID, accounts[0].ID)
 	require.Equal(t, due.ID, accounts[1].ID)
+	require.NotNil(t, accounts[1].LastUsedAt)
+	require.WithinDuration(t, activity.UTC(), accounts[1].LastUsedAt.UTC(), time.Second)
 	require.NotContains(t, accountIDs(accounts), uppercasePath.ID)
 	require.NotNil(t, accounts[0].Proxy)
 	require.Equal(t, proxy.ID, accounts[0].Proxy.ID)
 	require.Equal(t, proxy.URL(), accounts[0].Proxy.URL())
 }
 
-func TestListDueOllamaCloudUsageAccountsParsesRFC3339NanoAndFailsOpen(t *testing.T) {
+func TestListDueOllamaCloudUsageAccountsUsesGroupMaxLastUsedAndFailsOpen(t *testing.T) {
 	ctx := context.Background()
 	tx := testEntTx(t)
 	repo := newAccountRepositoryWithSQL(tx.Client(), tx, nil)
 	now := time.Date(2026, time.July, 22, 14, 0, 0, 0, time.UTC)
+	fetched := now.Add(-30 * time.Minute)
+	older := now.Add(-10 * time.Minute)
+	newer := now.Add(-2 * time.Minute)
 
-	create := func(name, nextRefreshAt string) *service.Account {
-		t.Helper()
-		return mustCreateAccount(t, tx.Client(), &service.Account{
-			Name: name, Platform: service.PlatformOpenAI, Type: service.AccountTypeAPIKey,
-			Credentials: map[string]any{"api_key": name, "base_url": "https://ollama.com"},
-			Extra: map[string]any{
-				service.OllamaCloudUsageSessionExtraKey:     "cipher:wos-session=fixture",
-				service.OllamaCloudUsageAutoRefreshExtraKey: true,
-				service.OllamaCloudUsageSnapshotExtraKey: map[string]any{
-					"status": service.OllamaCloudUsageStatusOK, "next_refresh_at": nextRefreshAt,
-				},
+	leader := mustCreateAccount(t, tx.Client(), &service.Account{
+		Name: "ollama-group-leader", Platform: service.PlatformOpenAI, Type: service.AccountTypeAPIKey,
+		Credentials: map[string]any{"api_key": "shared-key", "base_url": "https://ollama.com"},
+		Extra: map[string]any{
+			service.OllamaCloudUsageSessionExtraKey:     "cipher:wos-session=fixture",
+			service.OllamaCloudUsageAutoRefreshExtraKey: true,
+			service.OllamaCloudUsageSnapshotExtraKey: map[string]any{
+				"status":          service.OllamaCloudUsageStatusOK,
+				"fetched_at":      fetched.UTC().Format(time.RFC3339Nano),
+				"last_attempt_at": fetched.UTC().Format(time.RFC3339Nano),
+				"next_refresh_at": fetched.Add(time.Hour).UTC().Format(time.RFC3339Nano),
 			},
-		})
-	}
-
-	sevenDigitsOffset := create("ollama-nano-seven", "2026-07-22T11:00:00.1234567-02:00")
-	eightDigitsOffset := create("ollama-nano-eight", "2026-07-22T11:00:00.12345678+01:00")
-	nineDigitsZ := create("ollama-nano-nine", "2026-07-22T09:00:00.123456789Z")
-	invalidCalendar := create("ollama-nano-invalid", "2026-02-30T09:00:00.123456789Z")
-	future := create("ollama-nano-future", "2026-07-22T15:00:00.123456789Z")
+		},
+		LastUsedAt: &older,
+	})
+	_ = mustCreateAccount(t, tx.Client(), &service.Account{
+		Name: "ollama-group-sibling", Platform: service.PlatformAnthropic, Type: service.AccountTypeAPIKey,
+		Credentials: map[string]any{"api_key": "shared-key", "base_url": "https://www.ollama.com/v1"},
+		LastUsedAt:  &newer,
+	})
+	invalid := mustCreateAccount(t, tx.Client(), &service.Account{
+		Name: "ollama-invalid-snapshot", Platform: service.PlatformOpenAI, Type: service.AccountTypeAPIKey,
+		Credentials: map[string]any{"api_key": "invalid-key", "base_url": "https://ollama.com"},
+		Extra: map[string]any{
+			service.OllamaCloudUsageSessionExtraKey:     "cipher:wos-session=fixture",
+			service.OllamaCloudUsageAutoRefreshExtraKey: true,
+			service.OllamaCloudUsageSnapshotExtraKey: map[string]any{
+				"status": service.OllamaCloudUsageStatusOK, "fetched_at": "2026-02-30T09:00:00.123456789Z",
+			},
+		},
+	})
+	idle := mustCreateAccount(t, tx.Client(), &service.Account{
+		Name: "ollama-idle-ok", Platform: service.PlatformOpenAI, Type: service.AccountTypeAPIKey,
+		Credentials: map[string]any{"api_key": "idle-key", "base_url": "https://ollama.com"},
+		Extra: map[string]any{
+			service.OllamaCloudUsageSessionExtraKey:     "cipher:wos-session=fixture",
+			service.OllamaCloudUsageAutoRefreshExtraKey: true,
+			service.OllamaCloudUsageSnapshotExtraKey: map[string]any{
+				"status":          service.OllamaCloudUsageStatusOK,
+				"fetched_at":      fetched.UTC().Format(time.RFC3339Nano),
+				"last_attempt_at": fetched.UTC().Format(time.RFC3339Nano),
+				"next_refresh_at": fetched.Add(time.Hour).UTC().Format(time.RFC3339Nano),
+			},
+		},
+	})
 
 	accounts, err := repo.ListDueOllamaCloudUsageAccounts(ctx, now, 10)
 
 	require.NoError(t, err, "invalid stored values must not abort the query")
-	require.Equal(t, []int64{
-		invalidCalendar.ID,
-		nineDigitsZ.ID,
-		eightDigitsOffset.ID,
-		sevenDigitsOffset.ID,
-	}, accountIDs(accounts))
-	require.NotContains(t, accountIDs(accounts), future.ID)
+	ids := accountIDs(accounts)
+	require.Contains(t, ids, invalid.ID)
+	require.Contains(t, ids, leader.ID)
+	require.NotContains(t, ids, idle.ID)
+	for _, account := range accounts {
+		if account.ID == leader.ID {
+			require.NotNil(t, account.LastUsedAt)
+			require.WithinDuration(t, newer.UTC(), account.LastUsedAt.UTC(), time.Second,
+				"group MAX(last_used_at) must come from the sibling")
+		}
+	}
 }
 
 func TestLockAndMergeAccountProbeExtraCoalescesNullableOllamaGroupIdentity(t *testing.T) {
diff --git a/backend/internal/repository/account_repo_ollama_cloud_usage_test.go b/backend/internal/repository/account_repo_ollama_cloud_usage_test.go
index 07b67287ba..6adb819144 100644
--- a/backend/internal/repository/account_repo_ollama_cloud_usage_test.go
+++ b/backend/internal/repository/account_repo_ollama_cloud_usage_test.go
@@ -191,9 +191,9 @@ func TestListDueOllamaCloudUsageAccountsFiltersOrdersAndLimits(t *testing.T) {
 	t.Cleanup(func() { _ = db.Close() })
 	now := time.Date(2026, time.July, 22, 12, 0, 0, 0, time.UTC)
 	var capturedSQL string
-	mock.ExpectQuery("WITH candidates AS").
-		WithArgs(now, 20).
-		WillReturnRows(sqlmock.NewRows([]string{"id"}))
+	mock.ExpectQuery("WITH eligible AS").
+		WithArgs(20).
+		WillReturnRows(sqlmock.NewRows([]string{"id", "group_last_used_at"}))
 	repo := newAccountRepositoryWithSQL(nil, captureQuerySQL{db: db, captured: &capturedSQL}, nil)
 
 	accounts, err := repo.ListDueOllamaCloudUsageAccounts(context.Background(), now, 20)
@@ -209,10 +209,12 @@ func TestListDueOllamaCloudUsageAccountsFiltersOrdersAndLimits(t *testing.T) {
 		ollamaCloudBaseURLMatchesSQL("credentials ->> 'base_url'"),
 		"jsonb_typeof(extra -> 'ollama_cloud_usage_session') = 'string'",
 		`extra @> '{"ollama_cloud_usage_auto_refresh": true}'::jsonb`,
-		"parsed_next_refresh_at::timestamptz <= $1",
+		"MAX(last_used_at) AS group_last_used_at",
 		"PARTITION BY api_key",
 		"WHERE group_rank = 1",
-		"LIMIT $2",
+		"LIMIT $1",
+		"group_last_used_at > parsed_fetched_at::timestamptz",
+		"group_last_used_at > parsed_last_attempt_at::timestamptz",
 	} {
 		require.Contains(t, normalized, clause)
 	}
diff --git a/backend/internal/service/gateway_anthropic_passthrough.go b/backend/internal/service/gateway_anthropic_passthrough.go
index 2eb7a97e3b..8d7610b936 100644
--- a/backend/internal/service/gateway_anthropic_passthrough.go
+++ b/backend/internal/service/gateway_anthropic_passthrough.go
@@ -114,6 +114,9 @@ func (s *GatewayService) forwardAnthropicAPIKeyPassthroughWithInput(
 			if resp != nil && resp.Body != nil {
 				_ = resp.Body.Close()
 			}
+			if !errors.Is(err, context.Canceled) {
+				scheduleOllamaCloudUsageActivity(s.deferredService, account)
+			}
 			safeErr := sanitizeUpstreamErrorMessage(err.Error())
 			setOpsUpstreamError(c, 0, safeErr, "")
 			appendOpsUpstreamError(c, OpsUpstreamErrorEvent{
diff --git a/backend/internal/service/gateway_forward.go b/backend/internal/service/gateway_forward.go
index 2e6093d18e..3005d93fa1 100644
--- a/backend/internal/service/gateway_forward.go
+++ b/backend/internal/service/gateway_forward.go
@@ -369,6 +369,10 @@ func (s *GatewayService) Forward(ctx context.Context, c *gin.Context, account *A
 			if resp != nil && resp.Body != nil {
 				_ = resp.Body.Close()
 			}
+			// Transport attempt left local validation; count Ollama Cloud activity.
+			if !errors.Is(err, context.Canceled) {
+				scheduleOllamaCloudUsageActivity(s.deferredService, account)
+			}
 			// Ensure the client receives an error response (handlers assume Forward writes on non-failover errors).
 			safeErr := sanitizeUpstreamErrorMessage(err.Error())
 			setOpsUpstreamError(c, 0, safeErr, "")
diff --git a/backend/internal/service/gateway_upstream_response.go b/backend/internal/service/gateway_upstream_response.go
index 03f757dabc..a32c007e03 100644
--- a/backend/internal/service/gateway_upstream_response.go
+++ b/backend/internal/service/gateway_upstream_response.go
@@ -356,6 +356,8 @@ func (s *GatewayService) readUpstreamErrorBody(resp *http.Response) ([]byte, err
 }
 
 func (s *GatewayService) handleErrorResponse(ctx context.Context, resp *http.Response, c *gin.Context, account *Account, requestedModel ...string) (*ForwardResult, error) {
+	// Upstream returned a non-success HTTP status; count Ollama Cloud activity.
+	scheduleOllamaCloudUsageActivity(s.deferredService, account)
 	body, readErr := s.readUpstreamErrorBody(resp)
 	if readErr != nil {
 		// 读取失败时 body 可能被截断，错误分类会基于不完整数据；记录日志以便排查，
diff --git a/backend/internal/service/ollama_cloud_usage.go b/backend/internal/service/ollama_cloud_usage.go
index 2b50bab62c..1de9574a74 100644
--- a/backend/internal/service/ollama_cloud_usage.go
+++ b/backend/internal/service/ollama_cloud_usage.go
@@ -36,6 +36,9 @@ const (
 	ollamaCloudUsageDefaultIntervalMinutes = 60
 	ollamaCloudUsageMinIntervalMinutes     = 15
 	ollamaCloudUsageMaxIntervalMinutes     = 24 * 60
+	ollamaCloudUsageDefaultDebounceMinutes = 1
+	ollamaCloudUsageMinDebounceMinutes     = 1
+	ollamaCloudUsageMaxDebounceMinutes     = 60
 	ollamaCloudUsageCycleInterval          = time.Minute
 	ollamaCloudUsageManualRefreshInterval  = 30 * time.Second
 	ollamaCloudUsageRequestTimeout         = 15 * time.Second
@@ -76,10 +79,15 @@ const (
 	OllamaCloudUsageStatusFailed       = "failed"
 )
 
-// OllamaCloudUsageSettings controls the opt-in periodic refresh runner.
+// OllamaCloudUsageSettings controls the opt-in request-driven refresh runner.
+//
+// IntervalMinutes is the max-wait bound: when model requests keep arriving and
+// the trailing debounce keeps sliding, a refresh is forced after this long.
+// DebounceMinutes is the quiet period after the latest request in a group.
 type OllamaCloudUsageSettings struct {
 	Enabled         bool `json:"enabled"`
-	IntervalMinutes int  `json:"interval_minutes"`
+	IntervalMinutes int  `json:"interval_minutes"` // max wait while requests continue
+	DebounceMinutes int  `json:"debounce_minutes"` // trailing quiet period after last request
 }
 
 // OllamaCloudUsageWindow is a narrow, sanitized view of one official usage window.
@@ -114,6 +122,12 @@ type OllamaCloudUsageData struct {
 }
 
 // OllamaCloudUsageSnapshot is the only usage observation persisted in account extra.
+//
+// NextRefreshAt remains a persisted compatibility field. For status=ok it is a
+// max-wait horizon marker only; automatic success refreshes are driven by model
+// request activity (group last_used_at + debounce/max-wait), not by this field
+// alone. For failed/unauthorized snapshots it is the failure not-before time
+// (Retry-After / exponential backoff) and is enforced as max(activityDue, NextRefreshAt).
 type OllamaCloudUsageSnapshot struct {
 	Status        string                `json:"status"`
 	Data          *OllamaCloudUsageData `json:"data,omitempty"`
@@ -168,6 +182,9 @@ func (s *SettingService) GetOllamaCloudUsageSettings(ctx context.Context) (*Olla
 	if settings.IntervalMinutes == 0 {
 		settings.IntervalMinutes = defaults.IntervalMinutes
 	}
+	if settings.DebounceMinutes == 0 {
+		settings.DebounceMinutes = defaults.DebounceMinutes
+	}
 	normalizeOllamaCloudUsageSettings(&settings)
 	return &settings, nil
 }
@@ -185,6 +202,12 @@ func (s *SettingService) SetOllamaCloudUsageSettings(ctx context.Context, settin
 			fmt.Sprintf("interval_minutes must be between %d and %d", ollamaCloudUsageMinIntervalMinutes, ollamaCloudUsageMaxIntervalMinutes),
 		)
 	}
+	if settings.DebounceMinutes < ollamaCloudUsageMinDebounceMinutes || settings.DebounceMinutes > ollamaCloudUsageMaxDebounceMinutes {
+		return infraerrors.BadRequest(
+			"INVALID_OLLAMA_CLOUD_USAGE_DEBOUNCE",
+			fmt.Sprintf("debounce_minutes must be between %d and %d", ollamaCloudUsageMinDebounceMinutes, ollamaCloudUsageMaxDebounceMinutes),
+		)
+	}
 	normalizeOllamaCloudUsageSettings(settings)
 	data, err := json.Marshal(settings)
 	if err != nil {
@@ -194,7 +217,11 @@ func (s *SettingService) SetOllamaCloudUsageSettings(ctx context.Context, settin
 }
 
 func defaultOllamaCloudUsageSettings() *OllamaCloudUsageSettings {
-	return &OllamaCloudUsageSettings{Enabled: false, IntervalMinutes: ollamaCloudUsageDefaultIntervalMinutes}
+	return &OllamaCloudUsageSettings{
+		Enabled:         false,
+		IntervalMinutes: ollamaCloudUsageDefaultIntervalMinutes,
+		DebounceMinutes: ollamaCloudUsageDefaultDebounceMinutes,
+	}
 }
 
 func normalizeOllamaCloudUsageSettings(settings *OllamaCloudUsageSettings) {
@@ -204,6 +231,116 @@ func normalizeOllamaCloudUsageSettings(settings *OllamaCloudUsageSettings) {
 	if settings.IntervalMinutes > ollamaCloudUsageMaxIntervalMinutes {
 		settings.IntervalMinutes = ollamaCloudUsageMaxIntervalMinutes
 	}
+	if settings.DebounceMinutes <= 0 {
+		settings.DebounceMinutes = ollamaCloudUsageDefaultDebounceMinutes
+	}
+	if settings.DebounceMinutes < ollamaCloudUsageMinDebounceMinutes {
+		settings.DebounceMinutes = ollamaCloudUsageMinDebounceMinutes
+	}
+	if settings.DebounceMinutes > ollamaCloudUsageMaxDebounceMinutes {
+		settings.DebounceMinutes = ollamaCloudUsageMaxDebounceMinutes
+	}
+}
+
+func ollamaCloudUsageDurations(settings *OllamaCloudUsageSettings) (debounce, maxWait time.Duration) {
+	normalized := defaultOllamaCloudUsageSettings()
+	if settings != nil {
+		*normalized = *settings
+	}
+	normalizeOllamaCloudUsageSettings(normalized)
+	return time.Duration(normalized.DebounceMinutes) * time.Minute,
+		time.Duration(normalized.IntervalMinutes) * time.Minute
+}
+
+// ollamaCloudUsageIsAutoRefreshDue decides whether a configured auto-refresh
+// group should fetch now. groupLastUsedAt must be MAX(last_used_at) across the
+// exact api_key group so shared multi-platform accounts do not miss activity.
+//
+// Success: a request must be newer than fetched_at; dueAt = min(lastUsed+debounce, fetchedAt+maxWait).
+// Failure: a request must be newer than last_attempt_at; activity due uses the same min formula,
+// then dueAt = max(activityDue, next_refresh_at) so Retry-After / exponential backoff win.
+// Missing or invalid snapshots fail open to a first fetch.
+func ollamaCloudUsageIsAutoRefreshDue(
+	snapshot *OllamaCloudUsageSnapshot,
+	groupLastUsedAt *time.Time,
+	now time.Time,
+	debounce, maxWait time.Duration,
+) bool {
+	dueAt, ok := ollamaCloudUsageAutoRefreshDueAt(snapshot, groupLastUsedAt, debounce, maxWait)
+	if !ok {
+		return false
+	}
+	return !now.Before(dueAt)
+}
+
+func ollamaCloudUsageAutoRefreshDueAt(
+	snapshot *OllamaCloudUsageSnapshot,
+	groupLastUsedAt *time.Time,
+	debounce, maxWait time.Duration,
+) (time.Time, bool) {
+	if debounce <= 0 {
+		debounce = time.Duration(ollamaCloudUsageDefaultDebounceMinutes) * time.Minute
+	}
+	if maxWait <= 0 {
+		maxWait = time.Duration(ollamaCloudUsageDefaultIntervalMinutes) * time.Minute
+	}
+	if snapshot == nil {
+		return time.Time{}, true
+	}
+	switch snapshot.Status {
+	case OllamaCloudUsageStatusOK:
+		if snapshot.FetchedAt == nil || snapshot.FetchedAt.IsZero() {
+			return time.Time{}, true
+		}
+		fetchedAt := snapshot.FetchedAt.UTC()
+		if groupLastUsedAt == nil || !groupLastUsedAt.After(fetchedAt) {
+			return time.Time{}, false
+		}
+		lastUsed := groupLastUsedAt.UTC()
+		return minTime(lastUsed.Add(debounce), fetchedAt.Add(maxWait)), true
+	case OllamaCloudUsageStatusFailed, OllamaCloudUsageStatusUnauthorized:
+		if snapshot.LastAttemptAt.IsZero() {
+			return time.Time{}, true
+		}
+		lastAttempt := snapshot.LastAttemptAt.UTC()
+		if groupLastUsedAt == nil || !groupLastUsedAt.After(lastAttempt) {
+			return time.Time{}, false
+		}
+		lastUsed := groupLastUsedAt.UTC()
+		activityDue := minTime(lastUsed.Add(debounce), lastAttempt.Add(maxWait))
+		if !snapshot.NextRefreshAt.IsZero() && snapshot.NextRefreshAt.UTC().After(activityDue) {
+			return snapshot.NextRefreshAt.UTC(), true
+		}
+		return activityDue, true
+	default:
+		return time.Time{}, true
+	}
+}
+
+// maxOllamaCloudUsageGroupLastUsed returns the newest last_used_at among group members.
+func maxOllamaCloudUsageGroupLastUsed(accounts []Account) *time.Time {
+	var latest *time.Time
+	for i := range accounts {
+		candidate := accounts[i].LastUsedAt
+		if candidate == nil || candidate.IsZero() {
+			continue
+		}
+		if latest == nil || candidate.After(*latest) {
+			ts := candidate.UTC()
+			latest = &ts
+		}
+	}
+	return latest
+}
+
+// scheduleOllamaCloudUsageActivity records that an Ollama Cloud API-key account
+// actually attempted an upstream model request (including 429/5xx/transport errors).
+// Local auth/validation failures must not call this. DeferredService dedupes writes.
+func scheduleOllamaCloudUsageActivity(deferred *DeferredService, account *Account) {
+	if deferred == nil || account == nil || !IsOllamaCloudUsageAccount(account) {
+		return
+	}
+	deferred.ScheduleLastUsedUpdate(account.ID)
 }
 
 // OllamaCloudUsageService refreshes the official settings HTML without affecting routing state.
@@ -531,7 +668,7 @@ func (s *OllamaCloudUsageService) Refresh(ctx context.Context, accountID int64)
 	if err != nil {
 		return nil, err
 	}
-	if _, err := s.refreshAccount(ctx, accountID, settings.IntervalMinutes, false); err != nil {
+	if _, err := s.refreshAccount(ctx, accountID, settings, false); err != nil {
 		return nil, err
 	}
 	return s.GetState(ctx, accountID)
@@ -561,6 +698,7 @@ func (s *OllamaCloudUsageService) RunDue(ctx context.Context) error {
 		return ErrOllamaCloudUsageUnavailable
 	}
 	now := s.currentTime()
+	debounce, maxWait := ollamaCloudUsageDurations(settings)
 	accounts, err := writer.ListDueOllamaCloudUsageAccounts(ctx, now, ollamaCloudUsageMaxPerCycle)
 	if err != nil {
 		return fmt.Errorf("list due Ollama Cloud usage accounts: %w", err)
@@ -577,13 +715,15 @@ func (s *OllamaCloudUsageService) RunDue(ctx context.Context) error {
 			continue
 		}
 		seenGroups[fingerprint] = struct{}{}
-		if snapshot := decodeOllamaCloudUsageSnapshot(account.Extra); snapshot != nil && now.Before(snapshot.NextRefreshAt) {
+		snapshot := decodeOllamaCloudUsageSnapshot(account.Extra)
+		// ListDue stamps Account.LastUsedAt with the api_key group MAX(last_used_at).
+		if !ollamaCloudUsageIsAutoRefreshDue(snapshot, account.LastUsedAt, now, debounce, maxWait) {
 			continue
 		}
 		accountID := account.ID
 		expected := account
 		group.Go(func() error {
-			if _, refreshErr := s.refreshAccount(ctx, accountID, settings.IntervalMinutes, true); refreshErr != nil {
+			if _, refreshErr := s.refreshAccount(ctx, accountID, settings, true); refreshErr != nil {
 				if errors.Is(refreshErr, ErrOllamaCloudUsageIdentityChanged) {
 					if disableErr := writer.DisableOllamaCloudUsageAutoRefresh(ctx, &expected); disableErr != nil {
 						logger.LegacyPrintf("service.ollama_cloud_usage", "disable_auto_refresh_failed: account_id=%d err=%v", accountID, disableErr)
@@ -598,10 +738,15 @@ func (s *OllamaCloudUsageService) RunDue(ctx context.Context) error {
 	return group.Wait()
 }
 
-func (s *OllamaCloudUsageService) refreshAccount(ctx context.Context, accountID int64, intervalMinutes int, requireEnabled bool) (*OllamaCloudUsageSnapshot, error) {
+func (s *OllamaCloudUsageService) refreshAccount(ctx context.Context, accountID int64, settings *OllamaCloudUsageSettings, requireEnabled bool) (*OllamaCloudUsageSnapshot, error) {
 	if s == nil || s.accountRepo == nil {
 		return nil, ErrOllamaCloudUsageUnavailable
 	}
+	if settings == nil {
+		settings = defaultOllamaCloudUsageSettings()
+	}
+	intervalMinutes := settings.IntervalMinutes
+	debounce, maxWait := ollamaCloudUsageDurations(settings)
 	anchor, err := s.accountRepo.GetByID(ctx, accountID)
 	if err != nil {
 		return nil, err
@@ -650,7 +795,13 @@ func (s *OllamaCloudUsageService) refreshAccount(ctx context.Context, accountID
 			if !account.IsActive() || !ollamaCloudUsageAutoRefreshEnabled(account) {
 				return nil, nil
 			}
-			if snapshot := decodeOllamaCloudUsageSnapshot(account.Extra); snapshot != nil && s.currentTime().Before(snapshot.NextRefreshAt) {
+			groupLastUsed := account.LastUsedAt
+			if writer, ok := s.accountRepo.(ollamaCloudUsageRepository); ok {
+				if siblings, listErr := writer.ListOllamaCloudUsageGroupAccounts(ctx, []*Account{account}); listErr == nil {
+					groupLastUsed = maxOllamaCloudUsageGroupLastUsed(siblings)
+				}
+			}
+			if !ollamaCloudUsageIsAutoRefreshDue(decodeOllamaCloudUsageSnapshot(account.Extra), groupLastUsed, s.currentTime(), debounce, maxWait) {
 				return nil, nil
 			}
 		}
diff --git a/backend/internal/service/ollama_cloud_usage_test.go b/backend/internal/service/ollama_cloud_usage_test.go
index d949ce41fd..da73639a74 100644
--- a/backend/internal/service/ollama_cloud_usage_test.go
+++ b/backend/internal/service/ollama_cloud_usage_test.go
@@ -313,15 +313,87 @@ func TestOllamaCloudUsageSettingsDefaultOffAndValidation(t *testing.T) {
 	require.NoError(t, err)
 	require.False(t, settings.Enabled)
 	require.Equal(t, 60, settings.IntervalMinutes)
+	require.Equal(t, 1, settings.DebounceMinutes)
 
-	err = settingsService.SetOllamaCloudUsageSettings(context.Background(), &OllamaCloudUsageSettings{Enabled: true, IntervalMinutes: 14})
+	err = settingsService.SetOllamaCloudUsageSettings(context.Background(), &OllamaCloudUsageSettings{Enabled: true, IntervalMinutes: 14, DebounceMinutes: 1})
 	require.Error(t, err)
-	err = settingsService.SetOllamaCloudUsageSettings(context.Background(), &OllamaCloudUsageSettings{Enabled: true, IntervalMinutes: 90})
+	err = settingsService.SetOllamaCloudUsageSettings(context.Background(), &OllamaCloudUsageSettings{Enabled: true, IntervalMinutes: 90, DebounceMinutes: 0})
+	require.Error(t, err)
+	err = settingsService.SetOllamaCloudUsageSettings(context.Background(), &OllamaCloudUsageSettings{Enabled: true, IntervalMinutes: 90, DebounceMinutes: 61})
+	require.Error(t, err)
+	err = settingsService.SetOllamaCloudUsageSettings(context.Background(), &OllamaCloudUsageSettings{Enabled: true, IntervalMinutes: 90, DebounceMinutes: 2})
 	require.NoError(t, err)
 	settings, err = settingsService.GetOllamaCloudUsageSettings(context.Background())
 	require.NoError(t, err)
 	require.True(t, settings.Enabled)
 	require.Equal(t, 90, settings.IntervalMinutes)
+	require.Equal(t, 2, settings.DebounceMinutes)
+
+	// Legacy JSON without debounce_minutes defaults to 1.
+	repo.values[SettingKeyOllamaCloudUsageSettings] = `{"enabled":true,"interval_minutes":45}`
+	settings, err = settingsService.GetOllamaCloudUsageSettings(context.Background())
+	require.NoError(t, err)
+	require.Equal(t, 45, settings.IntervalMinutes)
+	require.Equal(t, 1, settings.DebounceMinutes)
+}
+
+func TestOllamaCloudUsageIsAutoRefreshDue(t *testing.T) {
+	debounce := time.Minute
+	maxWait := time.Hour
+	now := time.Date(2026, time.July, 25, 12, 0, 0, 0, time.UTC)
+	fetched := now.Add(-30 * time.Minute)
+	ptr := func(ts time.Time) *time.Time { return &ts }
+
+	require.True(t, ollamaCloudUsageIsAutoRefreshDue(nil, nil, now, debounce, maxWait), "missing snapshot first due")
+	require.True(t, ollamaCloudUsageIsAutoRefreshDue(&OllamaCloudUsageSnapshot{Status: "bogus"}, nil, now, debounce, maxWait), "invalid status first due")
+
+	okSnap := &OllamaCloudUsageSnapshot{
+		Status: OllamaCloudUsageStatusOK, FetchedAt: ptr(fetched),
+		LastAttemptAt: fetched, NextRefreshAt: fetched.Add(maxWait),
+	}
+	require.False(t, ollamaCloudUsageIsAutoRefreshDue(okSnap, nil, now, debounce, maxWait), "no request after success")
+	require.False(t, ollamaCloudUsageIsAutoRefreshDue(okSnap, ptr(fetched), now, debounce, maxWait), "request not after fetched_at")
+	require.False(t, ollamaCloudUsageIsAutoRefreshDue(okSnap, ptr(now.Add(-30*time.Second)), now, debounce, maxWait), "debounce not elapsed")
+	require.True(t, ollamaCloudUsageIsAutoRefreshDue(okSnap, ptr(now.Add(-time.Minute)), now, debounce, maxWait), "single request quiet for debounce")
+
+	// Continuous requests: last used is now, but max-wait from old fetch forces due.
+	oldFetched := now.Add(-2 * time.Hour)
+	oldSnap := &OllamaCloudUsageSnapshot{
+		Status: OllamaCloudUsageStatusOK, FetchedAt: ptr(oldFetched),
+		LastAttemptAt: oldFetched, NextRefreshAt: oldFetched.Add(maxWait),
+	}
+	require.True(t, ollamaCloudUsageIsAutoRefreshDue(oldSnap, ptr(now), now, debounce, maxWait), "max-wait forces due while requests continue")
+	// First request after a very old snapshot is immediately due because fetched+maxWait is past.
+	require.True(t, ollamaCloudUsageIsAutoRefreshDue(oldSnap, ptr(now.Add(-time.Second)), now, debounce, maxWait), "stale snapshot first request immediate")
+
+	failSnap := &OllamaCloudUsageSnapshot{
+		Status: OllamaCloudUsageStatusFailed, FetchedAt: ptr(fetched),
+		LastAttemptAt: now.Add(-10 * time.Minute), NextRefreshAt: now.Add(20 * time.Minute),
+	}
+	require.False(t, ollamaCloudUsageIsAutoRefreshDue(failSnap, nil, now, debounce, maxWait), "failure without new request")
+	require.False(t, ollamaCloudUsageIsAutoRefreshDue(failSnap, ptr(now.Add(-time.Minute)), now, debounce, maxWait), "failure blocked by backoff")
+	failSnap.NextRefreshAt = now.Add(-time.Second)
+	require.True(t, ollamaCloudUsageIsAutoRefreshDue(failSnap, ptr(now.Add(-time.Minute)), now, debounce, maxWait), "failure after backoff with new request")
+
+	require.True(t, ollamaCloudUsageIsAutoRefreshDue(&OllamaCloudUsageSnapshot{
+		Status: OllamaCloudUsageStatusOK, LastAttemptAt: now,
+	}, nil, now, debounce, maxWait), "ok without fetched_at fails open")
+}
+
+func TestScheduleOllamaCloudUsageActivityOnlyForOllama(t *testing.T) {
+	deferred := NewDeferredService(nil, nil, time.Second)
+	ollama := ollamaUsageAccount(1)
+	other := ollamaUsageAccount(2)
+	other.Credentials["base_url"] = "https://api.openai.com"
+
+	scheduleOllamaCloudUsageActivity(deferred, ollama)
+	scheduleOllamaCloudUsageActivity(deferred, other)
+	scheduleOllamaCloudUsageActivity(nil, ollama)
+
+	_, ok := deferred.lastUsedUpdates.Load(int64(1))
+	require.True(t, ok)
+	_, ok = deferred.lastUsedUpdates.Load(int64(2))
+	require.False(t, ok)
 }
 
 func TestIsOllamaCloudUsageAccountStrictOfficialHost(t *testing.T) {
diff --git a/backend/internal/service/openai_account_runtime_block_fastpath.go b/backend/internal/service/openai_account_runtime_block_fastpath.go
index c61f317859..ace8366a0e 100644
--- a/backend/internal/service/openai_account_runtime_block_fastpath.go
+++ b/backend/internal/service/openai_account_runtime_block_fastpath.go
@@ -51,6 +51,10 @@ func (s *OpenAIGatewayService) handleOpenAIAccountUpstreamError(ctx context.Cont
 	if account != nil && account.Platform == PlatformGrok && isGrokContentPolicyRejection(statusCode, responseBody) {
 		return false
 	}
+	// Any non-2xx upstream HTTP response means the model request was actually sent.
+	if s != nil {
+		scheduleOllamaCloudUsageActivity(s.deferredService, account)
+	}
 	stateCtx, cancel := openAIAccountStateContext(ctx)
 	defer cancel()
 
diff --git a/backend/internal/service/openai_upstream_transport_error.go b/backend/internal/service/openai_upstream_transport_error.go
index 2e2fa90f57..4093ca0b64 100644
--- a/backend/internal/service/openai_upstream_transport_error.go
+++ b/backend/internal/service/openai_upstream_transport_error.go
@@ -124,6 +124,11 @@ func (s *OpenAIGatewayService) handleOpenAIUpstreamTransportError(ctx context.Co
 		return err
 	}
 
+	// Transport attempt reached the network path; count as Ollama Cloud activity.
+	if s != nil {
+		scheduleOllamaCloudUsageActivity(s.deferredService, account)
+	}
+
 	if classifyOpenAITransportError(err).Persistent {
 		s.tempUnscheduleOpenAITransportError(ctx, account, safeErr)
 	}
diff --git a/frontend/src/api/__tests__/admin.accounts.ollamaCloudUsage.spec.ts b/frontend/src/api/__tests__/admin.accounts.ollamaCloudUsage.spec.ts
index 1777819cb1..e60643a695 100644
--- a/frontend/src/api/__tests__/admin.accounts.ollamaCloudUsage.spec.ts
+++ b/frontend/src/api/__tests__/admin.accounts.ollamaCloudUsage.spec.ts
@@ -38,7 +38,7 @@ describe('admin Ollama Cloud usage API', () => {
   })
 
   it('uses dedicated global settings endpoints', async () => {
-    const settings = { enabled: false, interval_minutes: 60 }
+    const settings = { enabled: false, interval_minutes: 60, debounce_minutes: 1 }
     get.mockResolvedValueOnce({ data: settings })
     put.mockResolvedValueOnce({ data: settings })
 
diff --git a/frontend/src/i18n/locales/en/admin/settings.ts b/frontend/src/i18n/locales/en/admin/settings.ts
index 45b9caca1e..8bc1a39cfc 100644
--- a/frontend/src/i18n/locales/en/admin/settings.ts
+++ b/frontend/src/i18n/locales/en/admin/settings.ts
@@ -335,11 +335,13 @@ export default {
       },
       ollamaCloudUsage: {
         title: 'Ollama Cloud Usage Refresh',
-        description: 'Periodically refresh official Ollama settings-page usage for individually opted-in accounts. Disabled by default.',
+        description: 'Refresh official Ollama settings-page usage driven by model requests for individually opted-in accounts. Disabled by default. Idle accounts are not polled.',
         enabled: 'Enable global automatic refresh',
-        enabledHint: 'Only accounts with a stored browser session and their own automatic refresh switch enabled are refreshed. Manual refresh remains available.',
-        intervalMinutes: 'Refresh interval (minutes)',
-        intervalHint: 'Range: 15–1440 minutes. Failures use bounded exponential backoff.',
+        enabledHint: 'Only accounts with a stored browser session and their own automatic refresh switch enabled are refreshed, and only after subsequent model requests. Manual refresh remains available.',
+        intervalMinutes: 'Max wait while requests continue (minutes)',
+        intervalHint: 'Range: 15–1440 minutes. When continuous requests keep sliding the debounce, force a refresh after this wait.',
+        debounceMinutes: 'Quiet period after last request (minutes)',
+        debounceHint: 'Range: 1–60 minutes. Refresh after the latest model request has been quiet for this long.',
         saved: 'Ollama Cloud usage refresh settings saved',
         saveFailed: 'Failed to save Ollama Cloud usage refresh settings'
       },
diff --git a/frontend/src/i18n/locales/zh/admin/settings.ts b/frontend/src/i18n/locales/zh/admin/settings.ts
index 4d7a2b23a1..dccf073c14 100644
--- a/frontend/src/i18n/locales/zh/admin/settings.ts
+++ b/frontend/src/i18n/locales/zh/admin/settings.ts
@@ -328,11 +328,13 @@ export default {
       },
       ollamaCloudUsage: {
         title: 'Ollama Cloud 用量刷新',
-        description: '定期刷新账号在 Ollama 官方设置页展示的用量；默认关闭。',
+        description: '在模型请求驱动下刷新账号在 Ollama 官方设置页展示的用量；默认关闭。无新请求时不会自动抓取。',
         enabled: '启用全局自动刷新',
-        enabledHint: '仅刷新已保存浏览器会话且账号自身也开启自动刷新的账号；手动刷新不受影响。',
-        intervalMinutes: '刷新周期（分钟）',
-        intervalHint: '范围 15–1440 分钟。失败后按有上限的指数退避重试。',
+        enabledHint: '仅刷新已保存浏览器会话且账号自身也开启自动刷新的账号；需有后续模型请求才会触发。手动刷新不受影响。',
+        intervalMinutes: '持续请求最长等待（分钟）',
+        intervalHint: '范围 15–1440 分钟。请求持续不断导致 debounce 一直后移时，最晚在此时间强制刷新。',
+        debounceMinutes: '请求安静等待（分钟）',
+        debounceHint: '范围 1–60 分钟。最后一次模型请求安静满此时长后再抓取用量。',
         saved: 'Ollama Cloud 用量刷新设置已保存',
         saveFailed: '保存 Ollama Cloud 用量刷新设置失败'
       },
diff --git a/frontend/src/types/index.ts b/frontend/src/types/index.ts
index 6198d8f832..98baf3c4ef 100644
--- a/frontend/src/types/index.ts
+++ b/frontend/src/types/index.ts
@@ -1023,7 +1023,10 @@ export interface OllamaCloudUsageState {
 
 export interface OllamaCloudUsageSettings {
   enabled: boolean
+  /** Max wait while model requests keep arriving (minutes). */
   interval_minutes: number
+  /** Trailing quiet period after the latest model request (minutes). */
+  debounce_minutes: number
 }
 
 export interface Account {
diff --git a/frontend/src/views/admin/SettingsView.vue b/frontend/src/views/admin/SettingsView.vue
index 3f50450b47..120a631740 100644
--- a/frontend/src/views/admin/SettingsView.vue
+++ b/frontend/src/views/admin/SettingsView.vue
@@ -4220,23 +4220,43 @@
                     data-testid="ollama-cloud-usage-global-enabled"
                   />
                 </div>
-                <div v-if="ollamaCloudUsageForm.enabled" class="border-t border-gray-100 pt-4 dark:border-dark-700">
-                  <label class="mb-2 block text-sm font-medium text-gray-700 dark:text-gray-300" for="ollama-cloud-usage-interval">
-                    {{ t("admin.settings.ollamaCloudUsage.intervalMinutes") }}
-                  </label>
-                  <input
-                    id="ollama-cloud-usage-interval"
-                    v-model.number="ollamaCloudUsageForm.interval_minutes"
-                    type="number"
-                    min="15"
-                    max="1440"
-                    class="input w-32"
-                    data-testid="ollama-cloud-usage-global-interval"
-                    @keydown.enter.prevent="saveOllamaCloudUsageSettings"
-                  />
-                  <p class="mt-1.5 text-xs text-gray-500 dark:text-gray-400">
-                    {{ t("admin.settings.ollamaCloudUsage.intervalHint") }}
-                  </p>
+                <div v-if="ollamaCloudUsageForm.enabled" class="space-y-4 border-t border-gray-100 pt-4 dark:border-dark-700">
+                  <div>
+                    <label class="mb-2 block text-sm font-medium text-gray-700 dark:text-gray-300" for="ollama-cloud-usage-debounce">
+                      {{ t("admin.settings.ollamaCloudUsage.debounceMinutes") }}
+                    </label>
+                    <input
+                      id="ollama-cloud-usage-debounce"
+                      v-model.number="ollamaCloudUsageForm.debounce_minutes"
+                      type="number"
+                      min="1"
+                      max="60"
+                      class="input w-32"
+                      data-testid="ollama-cloud-usage-global-debounce"
+                      @keydown.enter.prevent="saveOllamaCloudUsageSettings"
+                    />
+                    <p class="mt-1.5 text-xs text-gray-500 dark:text-gray-400">
+                      {{ t("admin.settings.ollamaCloudUsage.debounceHint") }}
+                    </p>
+                  </div>
+                  <div>
+                    <label class="mb-2 block text-sm font-medium text-gray-700 dark:text-gray-300" for="ollama-cloud-usage-interval">
+                      {{ t("admin.settings.ollamaCloudUsage.intervalMinutes") }}
+                    </label>
+                    <input
+                      id="ollama-cloud-usage-interval"
+                      v-model.number="ollamaCloudUsageForm.interval_minutes"
+                      type="number"
+                      min="15"
+                      max="1440"
+                      class="input w-32"
+                      data-testid="ollama-cloud-usage-global-interval"
+                      @keydown.enter.prevent="saveOllamaCloudUsageSettings"
+                    />
+                    <p class="mt-1.5 text-xs text-gray-500 dark:text-gray-400">
+                      {{ t("admin.settings.ollamaCloudUsage.intervalHint") }}
+                    </p>
+                  </div>
                 </div>
                 <div class="flex justify-end border-t border-gray-100 pt-4 dark:border-dark-700">
                   <button
@@ -7934,6 +7954,7 @@ const ollamaCloudUsageSaving = ref(false);
 const ollamaCloudUsageForm = reactive({
   enabled: false,
   interval_minutes: 60,
+  debounce_minutes: 1,
 });
 
 // Overload Cooldown (529) 状态
diff --git a/frontend/src/views/admin/__tests__/SettingsView.spec.ts b/frontend/src/views/admin/__tests__/SettingsView.spec.ts
index 5d888451df..76ecc6c6b5 100644
--- a/frontend/src/views/admin/__tests__/SettingsView.spec.ts
+++ b/frontend/src/views/admin/__tests__/SettingsView.spec.ts
@@ -50,6 +50,7 @@ const {
   getOllamaCloudUsageSettings: vi.fn().mockResolvedValue({
     enabled: false,
     interval_minutes: 60,
+    debounce_minutes: 1,
   }),
   updateOllamaCloudUsageSettings: vi.fn().mockImplementation(async (payload) => payload),
   getGroups: vi.fn(),
@@ -647,6 +648,7 @@ describe("admin SettingsView payment visible method controls", () => {
     getOllamaCloudUsageSettings.mockResolvedValue({
       enabled: false,
       interval_minutes: 60,
+      debounce_minutes: 1,
     });
     updateOllamaCloudUsageSettings.mockImplementation(async (payload) => payload);
     getGroups.mockResolvedValue([]);
@@ -1007,6 +1009,7 @@ describe("admin SettingsView payment visible method controls", () => {
     expect(card.find('[data-testid="ollama-cloud-usage-global-interval"]').exists()).toBe(false);
 
     await card.get('[data-testid="ollama-cloud-usage-global-enabled"]').setValue(true);
+    await card.get('[data-testid="ollama-cloud-usage-global-debounce"]').setValue(3);
     await card.get('[data-testid="ollama-cloud-usage-global-interval"]').setValue(90);
     await card.get('[data-testid="ollama-cloud-usage-global-save"]').trigger("click");
     await flushPromises();
@@ -1014,6 +1017,7 @@ describe("admin SettingsView payment visible method controls", () => {
     expect(updateOllamaCloudUsageSettings).toHaveBeenCalledWith({
       enabled: true,
       interval_minutes: 90,
+      debounce_minutes: 3,
     });
   });