diff --git a/open-sse/services/accountFallback.ts b/open-sse/services/accountFallback.ts
index a09b78924d..be048f69c9 100644
--- a/open-sse/services/accountFallback.ts
+++ b/open-sse/services/accountFallback.ts
@@ -184,6 +184,12 @@ export const CREDITS_EXHAUSTED_SIGNALS = [
   "out of credits",
   "payment required",
   "free tier of the model has been exhausted",
+  // #8631: narrower than a bare "has been exhausted" — that generic phrase also
+  // appears in Gemini's transient RPM/TPM 429 body ("Resource has been exhausted
+  // (e.g. check quota)."), which must stay RATE_LIMIT_EXCEEDED, not terminal.
+  // Anchoring on "tier" keeps free-tier depletion wording matched while excluding
+  // Gemini's "resource has been exhausted" rate-limit phrasing.
+  "tier has been exhausted",
   // #5239: providers (e.g. DeepSeek/GLM-style) return "Insufficient account balance"
   // on a depleted key. 402 is already terminalized by status, but catch non-402
   // out-of-credit bodies here too.
diff --git a/tests/unit/error-classifier.test.ts b/tests/unit/error-classifier.test.ts
index 4a13235779..afb6cd58c1 100644
--- a/tests/unit/error-classifier.test.ts
+++ b/tests/unit/error-classifier.test.ts
@@ -27,6 +27,11 @@ test("classifyProviderError: 400 + billing signal => QUOTA_EXHAUSTED", () => {
     error: { message: "insufficient_quota: exceeded your current quota" },
   });
   assert.equal(result, PROVIDER_ERROR_TYPES.QUOTA_EXHAUSTED);
+
+  const resultExhausted = classifyProviderError(400, {
+    error: { message: "The free tier of the model has been exhausted." },
+  });
+  assert.equal(resultExhausted, PROVIDER_ERROR_TYPES.QUOTA_EXHAUSTED);
 });
 
 test("classifyProviderError: Kimi billing-cycle 403 => QUOTA_EXHAUSTED", () => {