diff --git a/packages/ai/CHANGELOG.md b/packages/ai/CHANGELOG.md
index efc4386d8..7c1ed3af6 100644
--- a/packages/ai/CHANGELOG.md
+++ b/packages/ai/CHANGELOG.md
@@ -11,6 +11,10 @@
 - Added `durationMs` to `AssistantMessage`: `AssistantMessageEventStream` measures each response with a monotonic clock from the start of the request to its final message, for every API implementation, including direct calls. Deferred results fetched later stay untimed ([#10549](https://github.com/earendil-works/pi/issues/10549))
 - Added an optional `durationMs` to `ToolResultMessage` for the execution time of the tool ([#10549](https://github.com/earendil-works/pi/issues/10549))
 
+### Changed
+
+- The faux provider's prompt-cache usage estimate compares the previous and current prompt message by message and compares characters only from the first differing message; the usage numbers are unchanged.
+
 ### Fixed
 
 - Fixed `server_busy` and `servers are currently busy` provider errors ending the turn instead of being retried ([#10543](https://github.com/earendil-works/pi/issues/10543))
diff --git a/packages/ai/src/providers/faux.ts b/packages/ai/src/providers/faux.ts
index a38c18774..c3771cc2c 100644
--- a/packages/ai/src/providers/faux.ts
+++ b/packages/ai/src/providers/faux.ts
@@ -214,8 +214,23 @@ function messageToText(message: Message): string {
 	return toolResultToText(message);
 }
 
-function serializeContext(context: TranscriptContext): string {
-	return context.messages.map((message) => `${message.role}:${messageToText(message)}`).join("\n\n");
+/** Length of the prompt text that joins `messages` with blank lines. */
+function joinedLength(messages: readonly string[], count = messages.length): number {
+	let length = count > 0 ? (count - 1) * 2 : 0;
+	for (let index = 0; index < count; index++) length += messages[index]!.length;
+	return length;
+}
+
+/**
+ * Length of the common prefix of the two joined prompts. Equal messages are compared whole; characters are compared
+ * only from the first message that differs.
+ */
+function commonPromptPrefixLength(previous: readonly string[], current: readonly string[]): number {
+	let index = 0;
+	while (index < previous.length && index < current.length && previous[index] === current[index]) index++;
+	const rest = (messages: readonly string[]) =>
+		index === messages.length ? "" : (index > 0 ? "\n\n" : "") + messages.slice(index).join("\n\n");
+	return joinedLength(previous, index) + commonPrefixLength(rest(previous), rest(current));
 }
 
 function commonPrefixLength(a: string, b: string): number {
@@ -231,10 +246,12 @@ function withUsageEstimate(
 	message: AssistantMessage,
 	context: TranscriptContext,
 	options: StreamOptions | undefined,
-	promptCache: Map<string, string>,
+	promptCache: Map<string, readonly string[]>,
 ): AssistantMessage {
-	const promptText = serializeContext(context);
-	const promptTokens = estimateTokens(promptText);
+	// One text per message; the whole prompt joins them with blank lines.
+	const prompt = context.messages.map((message) => `${message.role}:${messageToText(message)}`);
+	const promptLength = joinedLength(prompt);
+	const promptTokens = Math.ceil(promptLength / 4);
 	const outputTokens = estimateTokens(assistantContentToText(message.content));
 	let input = promptTokens;
 	let cacheRead = 0;
@@ -244,14 +261,14 @@ function withUsageEstimate(
 	if (sessionId && options?.cacheRetention !== "none") {
 		const previousPrompt = promptCache.get(sessionId);
 		if (previousPrompt) {
-			const cachedChars = commonPrefixLength(previousPrompt, promptText);
-			cacheRead = estimateTokens(previousPrompt.slice(0, cachedChars));
-			cacheWrite = estimateTokens(promptText.slice(cachedChars));
+			const cachedChars = commonPromptPrefixLength(previousPrompt, prompt);
+			cacheRead = Math.ceil(cachedChars / 4);
+			cacheWrite = Math.ceil((promptLength - cachedChars) / 4);
 			input = Math.max(0, promptTokens - cacheRead);
 		} else {
 			cacheWrite = promptTokens;
 		}
-		promptCache.set(sessionId, promptText);
+		promptCache.set(sessionId, prompt);
 	}
 
 	return {
@@ -445,7 +462,7 @@ export function createFauxCore(options: RegisterFauxProviderOptions) {
 	let pendingResponses: FauxResponseStep[] = [];
 	const tokensPerSecond = options.tokensPerSecond;
 	const state: FauxProviderState = { callCount: 0, deferredFetchCount: 0, cancelledDeferred: [] };
-	const promptCache = new Map<string, string>();
+	const promptCache = new Map<string, readonly string[]>();
 	const deferredResponses = new Map<
 		string,
 		{
diff --git a/packages/ai/test/faux-provider.test.ts b/packages/ai/test/faux-provider.test.ts
index 557885f23..6f4178111 100644
--- a/packages/ai/test/faux-provider.test.ts
+++ b/packages/ai/test/faux-provider.test.ts
@@ -324,6 +324,27 @@ describe("faux provider", () => {
 		expect(second.usage.input + second.usage.cacheRead).toBeGreaterThan(second.usage.input);
 	});
 
+	it("counts cached characters up to the first difference in the joined prompt", async () => {
+		const registration = registerFauxProvider();
+		registrations.push(registration);
+		registration.setResponses([fauxAssistantMessage("a"), fauxAssistantMessage("b"), fauxAssistantMessage("c")]);
+		const options = { sessionId: "session-1", cacheRetention: "short" } as const;
+		const user = (content: string) => ({ role: "user" as const, content, timestamp: 1 });
+
+		// Prompt texts: "user:hello world" (16 characters), then 16 + 2 + "user:next" (9) = 27.
+		await complete(registration.getModel(), { messages: [user("hello world")] }, options);
+		const extended = await complete(
+			registration.getModel(),
+			{ messages: [user("hello world"), user("next")] },
+			options,
+		);
+		expect(extended.usage).toMatchObject({ input: 3, cacheRead: 4, cacheWrite: 3 });
+
+		// The first message now differs after "user:hello w" (12 characters): "user:hello wide" + 2 + 9 = 26.
+		const edited = await complete(registration.getModel(), { messages: [user("hello wide"), user("next")] }, options);
+		expect(edited.usage).toMatchObject({ input: 4, cacheRead: 3, cacheWrite: 4 });
+	});
+
 	it("does not simulate caching when cacheRetention is none", async () => {
 		const registration = registerFauxProvider();
 		registrations.push(registration);