diff --git a/extensions/posthog-analytics/lib/event-mappers.ts b/extensions/posthog-analytics/lib/event-mappers.ts
index 53e8693c48f..6f0783c3844 100644
--- a/extensions/posthog-analytics/lib/event-mappers.ts
+++ b/extensions/posthog-analytics/lib/event-mappers.ts
@@ -3,7 +3,13 @@ import type { TraceContextManager } from "./trace-context.js";
 import { readOrCreateAnonymousId, sanitizeMessages, sanitizeOutputChoices, stripSecrets } from "./privacy.js";
 
 /**
- * Extract actual token counts and cost from OpenClaw's per-message usage metadata.
+ * Extract token counts and cost from the LAST assistant message's usage metadata.
+ *
+ * Only the last assistant message is used so that each $ai_generation event
+ * reports the per-turn delta rather than a cumulative session total.  The
+ * previous implementation summed across ALL assistant messages, which meant
+ * multi-turn sessions emitted growing cumulative values on every turn and
+ * PostHog's sum() massively over-counted costs.
  */
 export function extractUsageFromMessages(messages: unknown): {
   inputTokens: number;
@@ -11,20 +17,23 @@ export function extractUsageFromMessages(messages: unknown): {
   totalCostUsd: number;
 } {
   if (!Array.isArray(messages)) return { inputTokens: 0, outputTokens: 0, totalCostUsd: 0 };
-  let inputTokens = 0;
-  let outputTokens = 0;
-  let totalCostUsd = 0;
+
+  let lastAssistantUsage: Record<string, unknown> | undefined;
   for (const msg of messages) {
     if (!msg || typeof msg !== "object") continue;
     const m = msg as Record<string, unknown>;
     if (m.role !== "assistant") continue;
     const usage = m.usage as Record<string, unknown> | undefined;
-    if (!usage) continue;
-    if (typeof usage.input === "number") inputTokens += usage.input;
-    if (typeof usage.output === "number") outputTokens += usage.output;
-    const cost = usage.cost as Record<string, unknown> | undefined;
-    if (cost && typeof cost.total === "number") totalCostUsd += cost.total;
+    if (usage) lastAssistantUsage = usage;
   }
+
+  if (!lastAssistantUsage) return { inputTokens: 0, outputTokens: 0, totalCostUsd: 0 };
+
+  const inputTokens = typeof lastAssistantUsage.input === "number" ? lastAssistantUsage.input : 0;
+  const outputTokens = typeof lastAssistantUsage.output === "number" ? lastAssistantUsage.output : 0;
+  const cost = lastAssistantUsage.cost as Record<string, unknown> | undefined;
+  const totalCostUsd = cost && typeof cost.total === "number" ? cost.total : 0;
+
   return { inputTokens, outputTokens, totalCostUsd };
 }
 
diff --git a/src/telemetry/event-mappers.test.ts b/src/telemetry/event-mappers.test.ts
index abc1304f9ad..e74ce564253 100644
--- a/src/telemetry/event-mappers.test.ts
+++ b/src/telemetry/event-mappers.test.ts
@@ -48,7 +48,7 @@ describe("extractToolNamesFromMessages", () => {
 });
 
 describe("extractUsageFromMessages", () => {
-  it("sums input/output tokens and cost across assistant messages", () => {
+  it("extracts usage from only the LAST assistant message (per-turn, not cumulative)", () => {
     const messages = [
       { role: "user", content: "hello" },
       { role: "assistant", content: "hi", usage: { input: 10, output: 50, cost: { total: 0.001 } } },
@@ -56,9 +56,9 @@ describe("extractUsageFromMessages", () => {
       { role: "assistant", content: "sure", usage: { input: 20, output: 100, cost: { total: 0.002 } } },
     ];
     const result = extractUsageFromMessages(messages);
-    expect(result.inputTokens).toBe(30);
-    expect(result.outputTokens).toBe(150);
-    expect(result.totalCostUsd).toBeCloseTo(0.003);
+    expect(result.inputTokens).toBe(20);
+    expect(result.outputTokens).toBe(100);
+    expect(result.totalCostUsd).toBeCloseTo(0.002);
   });
 
   it("skips non-assistant messages (user, tool, system)", () => {
@@ -83,6 +83,17 @@ describe("extractUsageFromMessages", () => {
   it("returns zeros for non-array input", () => {
     expect(extractUsageFromMessages(null)).toEqual({ inputTokens: 0, outputTokens: 0, totalCostUsd: 0 });
   });
+
+  it("uses the last assistant with usage even if a later assistant lacks it", () => {
+    const messages = [
+      { role: "assistant", content: "first", usage: { input: 50, output: 10, cost: { total: 0.01 } } },
+      { role: "assistant", content: "second (no usage)" },
+    ];
+    const result = extractUsageFromMessages(messages);
+    expect(result.inputTokens).toBe(50);
+    expect(result.outputTokens).toBe(10);
+    expect(result.totalCostUsd).toBe(0.01);
+  });
 });
 
 describe("normalizeOutputForPostHog", () => {