From 667fc3dd3608da6ff38410283d6d12d61f3fe087 Mon Sep 17 00:00:00 2001
From: David Brailovsky <davidbrai@gmail.com>
Date: Wed, 23 Sep 2026 17:16:32 +0200
Subject: [PATCH] fix(ai): price Vercel AI Gateway 1-hour cache writes
 correctly

Read the 1-hour cache-write count from streaming deltas when present, and add a regression test.

Fixes #9210
---
 packages/ai/CHANGELOG.md                      |  4 +++
 packages/ai/src/api/anthropic-messages.ts     |  7 ++++
 .../anthropic-cache-write-1h-cost.test.ts     | 35 +++++++++++++++++++
 3 files changed, 46 insertions(+)

diff --git a/packages/ai/CHANGELOG.md b/packages/ai/CHANGELOG.md
index 8ff552c5f..7ee21720f 100644
--- a/packages/ai/CHANGELOG.md
+++ b/packages/ai/CHANGELOG.md
@@ -6,6 +6,10 @@
 
 - Added `onProviderStreamEvent` to observe parsed provider stream events before normalization, including provider-specific fields not retained in assistant messages ([#9784](https://github.com/earendil-works/pi/issues/9784)).
 
+### Fixed
+
+- Fixed 1-hour Anthropic cache writes reported by Vercel AI Gateway in streaming deltas being priced at the 5-minute rate ([#9210](https://github.com/earendil-works/pi/issues/9210)).
+
 ## [0.87.1] - 2026-09-22
 
 ### Added
diff --git a/packages/ai/src/api/anthropic-messages.ts b/packages/ai/src/api/anthropic-messages.ts
index a935e01c5..9f227a245 100644
--- a/packages/ai/src/api/anthropic-messages.ts
+++ b/packages/ai/src/api/anthropic-messages.ts
@@ -776,6 +776,13 @@ export const stream: StreamFunction<"anthropic-messages", AnthropicOptions> = (
 						if (event.usage.cache_creation_input_tokens != null) {
 							output.usage.cacheWrite = event.usage.cache_creation_input_tokens;
 						}
+						// Vercel AI Gateway includes the TTL breakdown in deltas, though the SDK only types it on message_start.
+						const cacheCreation = (
+							event.usage as typeof event.usage & { cache_creation?: { ephemeral_1h_input_tokens?: number } }
+						).cache_creation;
+						if (cacheCreation?.ephemeral_1h_input_tokens != null) {
+							output.usage.cacheWrite1h = cacheCreation.ephemeral_1h_input_tokens;
+						}
 						// Anthropic reports reasoning tokens as a subset of output tokens.
 						const thinkingTokens = event.usage.output_tokens_details?.thinking_tokens;
 						if (thinkingTokens != null) {
diff --git a/packages/ai/test/anthropic-cache-write-1h-cost.test.ts b/packages/ai/test/anthropic-cache-write-1h-cost.test.ts
index 299c7dc50..dd4d82e2a 100644
--- a/packages/ai/test/anthropic-cache-write-1h-cost.test.ts
+++ b/packages/ai/test/anthropic-cache-write-1h-cost.test.ts
@@ -74,6 +74,41 @@ describe("Anthropic 1h cache write cost", () => {
 		expect(result.usage.cost.cacheWrite).toBeCloseTo(7.75, 10);
 	});
 
+	// Regression for #9210: Vercel AI Gateway sends cache usage in message_delta, not message_start.
+	it("prices 1h cache writes reported only in message_delta", async () => {
+		const model = getModel("vercel-ai-gateway", "anthropic/claude-haiku-4.5");
+		const response = createSseResponse([
+			{
+				event: "message_start",
+				data: JSON.stringify({
+					type: "message_start",
+					message: { id: "msg_test", usage: { input_tokens: 0, output_tokens: 0 } },
+				}),
+			},
+			{
+				event: "message_delta",
+				data: JSON.stringify({
+					type: "message_delta",
+					delta: { stop_reason: "end_turn" },
+					usage: {
+						input_tokens: 3,
+						output_tokens: 4,
+						cache_creation_input_tokens: 6535,
+						cache_creation: { ephemeral_5m_input_tokens: 0, ephemeral_1h_input_tokens: 6535 },
+					},
+				}),
+			},
+			{ event: "message_stop", data: JSON.stringify({ type: "message_stop" }) },
+		]);
+		const result = await streamAnthropic(model, context, {
+			client: createFakeAnthropicClient(response),
+		}).result();
+
+		expect(result.usage.cacheWrite).toBe(6535);
+		expect(result.usage.cacheWrite1h).toBe(6535);
+		expect(result.usage.cost.cacheWrite).toBeCloseTo((6535 * model.cost.input * 2) / 1_000_000, 10);
+	});
+
 	it("falls back to the 5m rate when no breakdown is reported", async () => {
 		const model = getModel("anthropic", "claude-opus-4-8");
 		const response = createSseResponse(eventsWithCacheCreation(undefined));