diff --git a/frontend/src/__tests__/recommendations.test.ts b/frontend/src/__tests__/recommendations.test.ts
index 70d2a0c33..a0a1bc07c 100644
--- a/frontend/src/__tests__/recommendations.test.ts
+++ b/frontend/src/__tests__/recommendations.test.ts
@@ -143,6 +143,12 @@ describe('Recommendations Module', () => {
     });
     (recsApi.refreshRecommendations as jest.Mock).mockResolvedValue({});
     mockShowToast.mockReturnValue({ dismiss: jest.fn() });
+
+    // Reset the auto-refresh in-flight guard AND the once-per-session
+    // latch so each test starts with a clean auto-refresh state. Without
+    // this, the first stale-load test latches autoRefreshAttempted and
+    // every later test in this block would silently skip the auto-refresh.
+    resetAutoRefreshInFlight();
   });
 
   afterEach(() => {
@@ -1456,6 +1462,48 @@ describe('Recommendations Module', () => {
       const inFlightOpts = inFlightCall![0] as { timeout?: number | null };
       expect(inFlightOpts.timeout).toBeNull();
     });
+
+    // Runaway-loop regression. Reproduces the production cost incident: the
+    // backend collect SUCCEEDS (POST /recommendations/refresh resolves) but the
+    // collector never advances last_collected_at (it fails at a DB error), so
+    // freshness stays stale forever. The success path calls onReload()
+    // (loadRecommendations by default), which re-enters triggerAutoRefreshIfStale.
+    // Pre-fix this recurses and re-fires the paid collect endpoint on every
+    // reload — observed at ~4-10 calls/sec in production. Post-fix the
+    // once-per-session latch caps it at exactly one collect across repeated
+    // loads / the post-refresh reload.
+    test('runaway loop — stale-that-never-clears fires the collect exactly once across reloads (#cost-risk)', async () => {
+      mockGetRecs();
+      // Freshness is ALWAYS stale (collector never advances it).
+      (recsApi.getRecommendationsFreshness as jest.Mock).mockResolvedValue({
+        last_collected_at: null,
+        last_collection_error: null,
+      });
+      // The refresh POST SUCCEEDS (real scenario: backend returns 202), which
+      // is what drives the success-path onReload() re-entry.
+      (recsApi.refreshRecommendations as jest.Mock).mockResolvedValue({});
+
+      // First page open: should auto-refresh once. The default onReload is
+      // loadRecommendations, so the success path re-enters this whole chain.
+      await loadRecommendations();
+      // Drain the .then(onReload) -> loadRecommendations -> triggerAutoRefreshIfStale
+      // microtask/promise chain several times; pre-fix each drain re-fires the
+      // collect, so a multi-iteration drain makes the runaway unmistakable.
+      for (let i = 0; i < 10; i++) {
+        await Promise.resolve();
+      }
+
+      // Even simulated subsequent reloads (filter change, tab re-entry) must
+      // NOT re-arm the auto-collect within the same session.
+      await loadRecommendations();
+      await loadRecommendations();
+      for (let i = 0; i < 10; i++) {
+        await Promise.resolve();
+      }
+
+      // Exactly one paid collect, regardless of how many reloads occurred.
+      expect(recsApi.refreshRecommendations).toHaveBeenCalledTimes(1);
+    });
   });
 
   describe('openPurchaseModal', () => {
diff --git a/frontend/src/recommendations.ts b/frontend/src/recommendations.ts
index 680556b80..e7c60611e 100644
--- a/frontend/src/recommendations.ts
+++ b/frontend/src/recommendations.ts
@@ -85,6 +85,22 @@ let lastRecommendationsSummary: RecommendationsSummary | null = null;
 // skips kicking off a second request (and duplicate toasts).
 let autoRefreshInFlight: Promise<void> | null = null;
 
+// Runaway-loop guard: latch so the stale-on-open auto-refresh fires AT MOST
+// once per page session. Without it, triggerAutoRefreshIfStale -> (stale)
+// -> startRecommendationsRefresh -> onReload() (loadRecommendations /
+// loadDashboard) re-enters triggerAutoRefreshIfStale, and when the backend
+// collection never advances last_collected_at (e.g. it fails at a DB error,
+// so freshness stays stale forever) the success path re-fires the collect
+// endlessly — observed at ~4-10 POST /recommendations/refresh per second in
+// production, each a paid Cost Explorer / pricing call once the collector
+// reaches AWS. autoRefreshInFlight only dedups *concurrent* refreshes; it is
+// cleared per cycle, so it cannot stop this sequential re-trigger. The latch
+// is intentionally NOT reset on success/failure: the auto-refresh is a
+// once-per-load convenience, and a hard page reload resets module state.
+// Explicit user actions (lookback change, manual refresh) call
+// startRecommendationsRefresh() directly and so are never gated by this latch.
+let autoRefreshAttempted = false;
+
 /**
  * Freshness budget for auto-refresh (#284). If the last successful collection
  * is older than this threshold (or there has never been one), loadRecommendations
@@ -159,6 +175,7 @@ export function resetExpandedCells(): void {
  */
 export function resetAutoRefreshInFlight(): void {
   autoRefreshInFlight = null;
+  autoRefreshAttempted = false;
 }
 
 /**
@@ -280,6 +297,16 @@ export function setupRecommendationsHandlers(): void {
 export async function triggerAutoRefreshIfStale(
   onReload: () => Promise<void> = loadRecommendations,
 ): Promise<void> {
+  // Runaway-loop guard: the stale-on-open auto-refresh runs at most once per
+  // page session. startRecommendationsRefresh's success path calls onReload()
+  // (loadRecommendations / loadDashboard), which re-enters this function; when
+  // the collector never advances last_collected_at the freshness stays stale
+  // and we would re-fire the (paid) collect endlessly. Latch before the
+  // freshness fetch so neither the GET nor the collect repeats. Explicit user
+  // refreshes bypass this by calling startRecommendationsRefresh() directly.
+  if (autoRefreshAttempted) return;
+  autoRefreshAttempted = true;
+
   let freshness;
   try {
     freshness = await getRecommendationsFreshness();