diff --git a/src/CodexCommands.ts b/src/CodexCommands.ts
index 321ac5da..0af237f6 100644
--- a/src/CodexCommands.ts
+++ b/src/CodexCommands.ts
@@ -476,9 +476,10 @@ export class CodexCommands {
         }
         const total = this.formatTokenCount(usage.totalTokens);
         const input = this.formatTokenCount(usage.inputTokens);
-        const cachedInput = this.formatTokenCount(usage.cachedInputTokens);
+        const cacheRead = this.formatTokenCount(usage.cachedInputTokens);
+        const cacheWrite = this.formatTokenCount(usage.cacheWriteInputTokens);
         const output = this.formatTokenCount(usage.outputTokens);
-        return `${total} total  (${input} input + ${cachedInput} cached input, ${output} output)`;
+        return `${total} total  (${input} input + ${cacheRead} cache read + ${cacheWrite} cache write, ${output} output)`;
     }
 
     private formatContextWindow(usage: TokenCount | null, contextWindow: number | null): string {
diff --git a/src/TokenCount.ts b/src/TokenCount.ts
index 1f7cf0b5..83e4d671 100644
--- a/src/TokenCount.ts
+++ b/src/TokenCount.ts
@@ -5,16 +5,18 @@ import type {TokenUsageBreakdown} from "./app-server/v2";
  * Token usage information for a turn.
  * This interface decouples our API from Codex's internal types.
  *
- * [totalTokens]: total number of tokens used (the sum of all other fields)
- * [inputTokens]: number of non-cached input tokens
- * [cachedInputTokens]: number of cached input tokens
+ * [totalTokens]: ordinary input + cache reads + cache writes + output
+ * [inputTokens]: number of ordinary input tokens
+ * [cachedInputTokens]: number of cache-read input tokens
+ * [cacheWriteInputTokens]: number of cache-write input tokens
  * [outputTokens]: number of output tokens (including reasoning output tokens)
- * [reasoningOutputTokens]: number of reasoning output tokens
+ * [reasoningOutputTokens]: reasoning subset of output tokens
  */
 export interface TokenCount {
     totalTokens: number;
     inputTokens: number;
     cachedInputTokens: number;
+    cacheWriteInputTokens: number;
     outputTokens: number;
     reasoningOutputTokens: number;
 }
@@ -22,14 +24,16 @@ export interface TokenCount {
 /**
  * Maps Codex's TokenUsageBreakdown to our TokenCount interface.
  * This explicit mapping ensures compile-time errors if Codex changes their types.
- * Note: Codex includes cached input tokens in the input token count, so they are subtracted here.
+ * Note: Codex includes cache reads and writes in the input token count, so they are subtracted here.
  */
 export function toTokenCount(usage: TokenUsageBreakdown): TokenCount {
+    const cacheWriteInputTokens = usage.cacheWriteInputTokens ?? 0;
 
     return {
         totalTokens: usage.totalTokens,
-        inputTokens: usage.inputTokens - usage.cachedInputTokens,
+        inputTokens: usage.inputTokens - usage.cachedInputTokens - cacheWriteInputTokens,
         cachedInputTokens: usage.cachedInputTokens,
+        cacheWriteInputTokens:  cacheWriteInputTokens,
         outputTokens: usage.outputTokens,
         reasoningOutputTokens: usage.reasoningOutputTokens,
     };
@@ -37,14 +41,15 @@ export function toTokenCount(usage: TokenUsageBreakdown): TokenCount {
 
 /**
  * Maps our per-turn token breakdown to ACP PromptResponse usage fields.
- * Cached input tokens are reported as ACP cache reads, and reasoning output
- * tokens are exposed through ACP's thoughtTokens field.
+ * Cache reads and writes are reported through their corresponding ACP fields,
+ * and reasoning output tokens are exposed through ACP's thoughtTokens field.
  */
 export function toPromptUsage(tokenCount: TokenCount): Usage {
     return {
         totalTokens: tokenCount.totalTokens,
         inputTokens: tokenCount.inputTokens,
         cachedReadTokens: tokenCount.cachedInputTokens,
+        cachedWriteTokens: tokenCount.cacheWriteInputTokens,
         outputTokens: tokenCount.outputTokens,
         thoughtTokens: tokenCount.reasoningOutputTokens,
     };
diff --git a/src/__tests__/CodexACPAgent/CodexAcpClient.test.ts b/src/__tests__/CodexACPAgent/CodexAcpClient.test.ts
index fe51f160..3739ceca 100644
--- a/src/__tests__/CodexACPAgent/CodexAcpClient.test.ts
+++ b/src/__tests__/CodexACPAgent/CodexAcpClient.test.ts
@@ -1624,8 +1624,17 @@ describe('ACP server test', { timeout: 40_000 }, () => {
         const {mockFixture, sessionState} = setupPromptFixture({
             lastTokenUsage: {
                 totalTokens: 41_500,
-                inputTokens: 40_000,
+                inputTokens: 39_000,
+                cachedInputTokens: 1_000,
+                cacheWriteInputTokens: 1_000,
+                outputTokens: 500,
+                reasoningOutputTokens: 100,
+            },
+            totalTokenUsage: {
+                totalTokens: 41_500,
+                inputTokens: 39_000,
                 cachedInputTokens: 1_000,
+                cacheWriteInputTokens: 1_000,
                 outputTokens: 500,
                 reasoningOutputTokens: 100,
             },
@@ -1641,6 +1650,9 @@ describe('ACP server test', { timeout: 40_000 }, () => {
         expect(mockFixture.getAcpConnectionDump([])).toContain(
             "**Context window:** 16% used (41.5K used / 258.4K)",
         );
+        expect(mockFixture.getAcpConnectionDump([])).toContain(
+            "**Token usage:** 41.5K total  (39.0K input + 1.0K cache read + 1.0K cache write, 500 output)",
+        );
     });
 
     it('resets the previous context usage before a model turn', async () => {
@@ -1649,6 +1661,7 @@ describe('ACP server test', { timeout: 40_000 }, () => {
                 totalTokens: 41_500,
                 inputTokens: 40_000,
                 cachedInputTokens: 1_000,
+                cacheWriteInputTokens: 0,
                 outputTokens: 500,
                 reasoningOutputTokens: 100,
             },
diff --git a/src/__tests__/CodexACPAgent/data/token-usage-cancelled.json b/src/__tests__/CodexACPAgent/data/token-usage-cancelled.json
index 426011ab..f620b831 100644
--- a/src/__tests__/CodexACPAgent/data/token-usage-cancelled.json
+++ b/src/__tests__/CodexACPAgent/data/token-usage-cancelled.json
@@ -4,6 +4,7 @@
     "totalTokens": 1500,
     "inputTokens": 1200,
     "cachedReadTokens": 0,
+    "cachedWriteTokens": 0,
     "outputTokens": 300,
     "thoughtTokens": 0
   },
@@ -13,6 +14,7 @@
         "totalTokens": 1500,
         "inputTokens": 1200,
         "cachedInputTokens": 0,
+        "cacheWriteInputTokens": 0,
         "outputTokens": 300,
         "reasoningOutputTokens": 0
       },
@@ -23,6 +25,7 @@
             "totalTokens": 1500,
             "inputTokens": 1200,
             "cachedInputTokens": 0,
+            "cacheWriteInputTokens": 0,
             "outputTokens": 300,
             "reasoningOutputTokens": 0
           }
diff --git a/src/__tests__/CodexACPAgent/data/token-usage-end-turn.json b/src/__tests__/CodexACPAgent/data/token-usage-end-turn.json
index 60ba2e2e..80b203b2 100644
--- a/src/__tests__/CodexACPAgent/data/token-usage-end-turn.json
+++ b/src/__tests__/CodexACPAgent/data/token-usage-end-turn.json
@@ -2,18 +2,20 @@
   "stopReason": "end_turn",
   "usage": {
     "totalTokens": 2500,
-    "inputTokens": 1500,
+    "inputTokens": 1250,
     "cachedReadTokens": 500,
-    "outputTokens": 450,
+    "cachedWriteTokens": 250,
+    "outputTokens": 500,
     "thoughtTokens": 50
   },
   "_meta": {
     "quota": {
       "token_count": {
         "totalTokens": 2500,
-        "inputTokens": 1500,
+        "inputTokens": 1250,
         "cachedInputTokens": 500,
-        "outputTokens": 450,
+        "cacheWriteInputTokens": 250,
+        "outputTokens": 500,
         "reasoningOutputTokens": 50
       },
       "model_usage": [
@@ -21,9 +23,10 @@
           "model": "model-id",
           "token_count": {
             "totalTokens": 2500,
-            "inputTokens": 1500,
+            "inputTokens": 1250,
             "cachedInputTokens": 500,
-            "outputTokens": 450,
+            "cacheWriteInputTokens": 250,
+            "outputTokens": 500,
             "reasoningOutputTokens": 50
           }
         }
diff --git a/src/__tests__/CodexACPAgent/data/token-usage-multiple-updates.json b/src/__tests__/CodexACPAgent/data/token-usage-multiple-updates.json
index aebc1783..d8c346db 100644
--- a/src/__tests__/CodexACPAgent/data/token-usage-multiple-updates.json
+++ b/src/__tests__/CodexACPAgent/data/token-usage-multiple-updates.json
@@ -4,7 +4,8 @@
     "totalTokens": 1500,
     "inputTokens": 700,
     "cachedReadTokens": 500,
-    "outputTokens": 200,
+    "cachedWriteTokens": 0,
+    "outputTokens": 300,
     "thoughtTokens": 100
   },
   "_meta": {
@@ -13,7 +14,8 @@
         "totalTokens": 1500,
         "inputTokens": 700,
         "cachedInputTokens": 500,
-        "outputTokens": 200,
+        "cacheWriteInputTokens": 0,
+        "outputTokens": 300,
         "reasoningOutputTokens": 100
       },
       "model_usage": [
@@ -23,7 +25,8 @@
             "totalTokens": 1500,
             "inputTokens": 700,
             "cachedInputTokens": 500,
-            "outputTokens": 200,
+            "cacheWriteInputTokens": 0,
+            "outputTokens": 300,
             "reasoningOutputTokens": 100
           }
         }
diff --git a/src/__tests__/CodexACPAgent/token-usage-events.test.ts b/src/__tests__/CodexACPAgent/token-usage-events.test.ts
index dc49c1fc..b47860a0 100644
--- a/src/__tests__/CodexACPAgent/token-usage-events.test.ts
+++ b/src/__tests__/CodexACPAgent/token-usage-events.test.ts
@@ -60,16 +60,16 @@ describe('Token Usage Events', () => {
                     totalTokens: 5000,
                     inputTokens: 4000,
                     cachedInputTokens: 1000,
-                    cacheWriteInputTokens: 0,
-                    outputTokens: 900,
+                    cacheWriteInputTokens: 500,
+                    outputTokens: 1000,
                     reasoningOutputTokens: 100,
                 },
                 last: {
                     totalTokens: 2500,
                     inputTokens: 2000,
                     cachedInputTokens: 500,
-                    cacheWriteInputTokens: 0,
-                    outputTokens: 450,
+                    cacheWriteInputTokens: 250,
+                    outputTokens: 500,
                     reasoningOutputTokens: 50,
                 },
                 modelContextWindow: 128000,
@@ -87,24 +87,23 @@ describe('Token Usage Events', () => {
             );
         });
 
-        it('should include token_count in PromptResponse on cancelled', async () => {
+        it('should include token_count on cancelled with legacy token usage', async () => {
             const tokenUsageNotification = createTokenUsageNotification(sessionId, {
+                // Older Codex binaries can omit cacheWriteInputTokens at runtime.
                 total: {
                     totalTokens: 3000,
                     inputTokens: 2500,
                     cachedInputTokens: 0,
-                    cacheWriteInputTokens: 0,
                     outputTokens: 500,
                     reasoningOutputTokens: 0,
-                },
+                } as TokenUsageBreakdown,
                 last: {
                     totalTokens: 1500,
                     inputTokens: 1200,
                     cachedInputTokens: 0,
-                    cacheWriteInputTokens: 0,
                     outputTokens: 300,
                     reasoningOutputTokens: 0,
-                },
+                } as TokenUsageBreakdown,
                 modelContextWindow: 128000,
             });
 
@@ -146,8 +145,8 @@ describe('Token Usage Events', () => {
                     modelContextWindow: 128000,
                 }),
                 createTokenUsageNotification(sessionId, {
-                    total: { totalTokens: 3500, inputTokens: 2800, cachedInputTokens: 500, cacheWriteInputTokens: 0, outputTokens: 600, reasoningOutputTokens: 100 },
-                    last: { totalTokens: 1500, inputTokens: 1200, cachedInputTokens: 500, cacheWriteInputTokens: 0, outputTokens: 200, reasoningOutputTokens: 100 },
+                    total: { totalTokens: 3500, inputTokens: 2800, cachedInputTokens: 500, cacheWriteInputTokens: 0, outputTokens: 700, reasoningOutputTokens: 100 },
+                    last: { totalTokens: 1500, inputTokens: 1200, cachedInputTokens: 500, cacheWriteInputTokens: 0, outputTokens: 300, reasoningOutputTokens: 100 },
                     modelContextWindow: 128000,
                 }),
             ];
@@ -201,16 +200,16 @@ describe('Token Usage Events', () => {
                         totalTokens: 5000,
                         inputTokens: 4000,
                         cachedInputTokens: 1000,
-                        cacheWriteInputTokens: 0,
-                        outputTokens: 900,
+                        cacheWriteInputTokens: 500,
+                        outputTokens: 1000,
                         reasoningOutputTokens: 100,
                     },
                     last: {
                         totalTokens: 2500,
                         inputTokens: 2000,
                         cachedInputTokens: 500,
-                        cacheWriteInputTokens: 0,
-                        outputTokens: 450,
+                        cacheWriteInputTokens: 250,
+                        outputTokens: 500,
                         reasoningOutputTokens: 50,
                     },
                     modelContextWindow: 128000,
@@ -233,8 +232,8 @@ describe('Token Usage Events', () => {
                     modelContextWindow: 128000,
                 }),
                 createTokenUsageNotification(sessionId, {
-                    total: { totalTokens: 3500, inputTokens: 2800, cachedInputTokens: 500, cacheWriteInputTokens: 0, outputTokens: 600, reasoningOutputTokens: 100 },
-                    last: { totalTokens: 1500, inputTokens: 1200, cachedInputTokens: 500, cacheWriteInputTokens: 0, outputTokens: 200, reasoningOutputTokens: 100 },
+                    total: { totalTokens: 3500, inputTokens: 2800, cachedInputTokens: 500, cacheWriteInputTokens: 0, outputTokens: 700, reasoningOutputTokens: 100 },
+                    last: { totalTokens: 1500, inputTokens: 1200, cachedInputTokens: 500, cacheWriteInputTokens: 0, outputTokens: 300, reasoningOutputTokens: 100 },
                     modelContextWindow: 128000,
                 }),
             ])();