diff --git a/packages/ai/src/auth-storage.ts b/packages/ai/src/auth-storage.ts
index 2d6af2212..33b4a006b 100644
--- a/packages/ai/src/auth-storage.ts
+++ b/packages/ai/src/auth-storage.ts
@@ -3203,13 +3203,12 @@ export class AuthStorage {
 			};
 		}
 
-		// Header-only reports are last-good hints, not completed usage fetches.
-		// Keep them durable but stale so the next poll probes the provider, while
-		// preserving any active failure backoff installed by #fetchUsageCached.
-		const expiresAt =
-			merged.metadata?.source === "ratelimit-headers"
-				? Math.max(priorEntry?.expiresAt ?? now - 1, now - 1)
-				: now + USAGE_REPORT_TTL_MS;
+		// Header ingestion merges values but never extends a cache entry's lifetime.
+		// Preserve the existing expiry (including active failure cooldowns) so full
+		// reports refetch on their original 5-minute schedule and full-payload-only
+		// rows such as extra usage stay current; headers only refresh window rows
+		// between fetches. A newly minted header-only report is durable but stale.
+		const expiresAt = Math.max(priorEntry?.expiresAt ?? now - 1, now - 1);
 		this.#usageCache.set(cacheKey, { value: merged, expiresAt });
 		this.#usageHeaderIngestAt.set(cacheKey, now);
 		return true;
diff --git a/packages/ai/test/auth-storage-usage-cache.test.ts b/packages/ai/test/auth-storage-usage-cache.test.ts
index 48bb8910d..10f13671f 100644
--- a/packages/ai/test/auth-storage-usage-cache.test.ts
+++ b/packages/ai/test/auth-storage-usage-cache.test.ts
@@ -492,6 +492,74 @@ describe("AuthStorage usage cache: header ingestion", () => {
 		expect(calls).toBe(2);
 	});
 
+	it("does not let header ingestion slide the full-report refresh deadline", async () => {
+		const start = Date.now();
+		const now = vi.spyOn(Date, "now").mockReturnValue(start);
+		vi.spyOn(Math, "random").mockReturnValue(0.5);
+		const makeFullReport = (extraUsed: number): UsageReport => {
+			const baseReport = makeTieredReport("a@example.com");
+			return {
+				...baseReport,
+				limits: [
+					...baseReport.limits,
+					{
+						id: "anthropic:extra",
+						label: "Claude Extra Usage",
+						scope: { provider: "anthropic", windowId: "extra" },
+						amount: {
+							used: extraUsed,
+							limit: 100,
+							usedFraction: extraUsed / 100,
+							unit: "usd",
+						},
+						status: "ok",
+					},
+				],
+				raw: { extra_usage: { used: extraUsed * 100, limit: 10_000 } },
+			};
+		};
+		const firstFullReport = makeFullReport(12.34);
+		const secondFullReport = makeFullReport(56.78);
+		const fetchSpy = vi
+			.spyOn(claudeUsage.claudeUsageProvider, "fetchUsage")
+			.mockResolvedValueOnce(firstFullReport)
+			.mockResolvedValue(secondFullReport);
+
+		const initialReport = requireAnthropicReport(await storage.fetchUsageReports());
+		expect(fetchSpy).toHaveBeenCalledTimes(1);
+		expect(requireLimit(initialReport, "anthropic:extra").amount.used).toBe(12.34);
+		expect(await storage.getApiKey("anthropic", "sliding-session")).toBe("oat-1");
+
+		now.mockReturnValue(start + 60_000);
+		expect(
+			storage.ingestUsageHeaders("anthropic", usageHeaders("0.05", "0.6"), {
+				sessionId: "sliding-session",
+			}),
+		).toBe(true);
+		now.mockReturnValue(start + 120_000);
+		expect(
+			storage.ingestUsageHeaders("anthropic", usageHeaders("0.06", "0.61"), {
+				sessionId: "sliding-session",
+			}),
+		).toBe(true);
+		now.mockReturnValue(start + 240_000);
+		expect(
+			storage.ingestUsageHeaders("anthropic", usageHeaders("0.07", "0.62"), {
+				sessionId: "sliding-session",
+			}),
+		).toBe(true);
+
+		now.mockReturnValue(start + 299_999);
+		const beforeDeadline = requireAnthropicReport(await storage.fetchUsageReports());
+		expect(fetchSpy).toHaveBeenCalledTimes(1);
+		expect(requireLimit(beforeDeadline, "anthropic:extra").amount.used).toBe(12.34);
+
+		now.mockReturnValue(start + 376_000);
+		const refreshed = requireAnthropicReport(await storage.fetchUsageReports());
+		expect(fetchSpy).toHaveBeenCalledTimes(2);
+		expect(requireLimit(refreshed, "anthropic:extra").amount.used).toBe(56.78);
+	});
+
 	it("merges header umbrella windows onto the last real report and preserves tier limits", async () => {
 		const realReport = makeTieredReport("a@example.com");
 		let calls = 0;