diff --git a/docs/images/token-output-speed.png b/docs/images/token-output-speed.png
new file mode 100644
index 000000000..a45f34fd2
Binary files /dev/null and b/docs/images/token-output-speed.png differ
diff --git a/src/components/chat/message-input.tsx b/src/components/chat/message-input.tsx
index 2d88beee4..adfa3d20b 100644
--- a/src/components/chat/message-input.tsx
+++ b/src/components/chat/message-input.tsx
@@ -127,6 +127,7 @@ import {
 } from "@/components/chat/conversation-context-bar"
 import { ComposerContextUsage } from "@/components/chat/composer-context-usage"
 import { ComposerConnectionStatus } from "@/components/chat/composer-connection-status"
+import { TokenOutputSpeed } from "@/components/chat/token-output-speed"
 import { InlineModeSelector } from "@/components/chat/mode-selector"
 import { InlineSessionConfigSelector } from "@/components/chat/session-config-selector"
 import { ModelOptionPicker } from "@/components/chat/model-option-picker"
@@ -3872,6 +3873,7 @@ export function MessageInput({
                 send button's right edge in the action bar above — no centring
                 slot, which would inset the narrow icon and break the alignment. */}
             <div className="flex shrink-0 items-center gap-3 pr-px">
+              <TokenOutputSpeed tabId={attachmentTabId ?? null} />
               <ComposerContextUsage tabId={attachmentTabId ?? null} />
               <ComposerConnectionStatus tabId={attachmentTabId ?? null} />
             </div>
diff --git a/src/components/chat/token-output-speed.test.tsx b/src/components/chat/token-output-speed.test.tsx
new file mode 100644
index 000000000..0c2222352
--- /dev/null
+++ b/src/components/chat/token-output-speed.test.tsx
@@ -0,0 +1,128 @@
+import { act, render, screen } from "@testing-library/react"
+import { NextIntlClientProvider } from "next-intl"
+import { afterEach, describe, expect, it, vi } from "vitest"
+
+import { TokenOutputSpeed } from "./token-output-speed"
+import enMessages from "@/i18n/messages/en.json"
+import type {
+  ConnectionState,
+  LiveContentBlock,
+} from "@/contexts/acp-connections-context"
+
+const mock = vi.hoisted(() => {
+  const state: {
+    conn: ConnectionState | undefined
+    listener: (() => void) | null
+  } = {
+    conn: undefined,
+    listener: null,
+  }
+  return {
+    subscribeKey: (_key: string, cb: () => void) => {
+      state.listener = cb
+      return () => {
+        if (state.listener === cb) state.listener = null
+      }
+    },
+    getConnection: () => state.conn,
+    emit: () => state.listener?.(),
+    setConn: (conn: ConnectionState | undefined) => {
+      state.conn = conn
+    },
+  }
+})
+
+vi.mock("@/contexts/acp-connections-context", () => ({
+  useConnectionStore: () => ({
+    subscribeKey: mock.subscribeKey,
+    getConnection: mock.getConnection,
+  }),
+}))
+
+function conn(blocks: LiveContentBlock[]): ConnectionState {
+  return {
+    connectionId: "c1",
+    contextKey: "tab1",
+    agentType: "codex",
+    workingDir: null,
+    status: "prompting",
+    liveMessage: {
+      id: "live-1",
+      role: "assistant",
+      content: blocks,
+      startedAt: 0,
+    },
+  } as unknown as ConnectionState
+}
+
+function renderBadge() {
+  return render(
+    <NextIntlClientProvider messages={enMessages} locale="en">
+      <TokenOutputSpeed tabId="tab1" />
+    </NextIntlClientProvider>
+  )
+}
+
+let fakeNow = 0
+
+afterEach(() => {
+  vi.restoreAllMocks()
+  fakeNow = 0
+})
+
+describe("TokenOutputSpeed", () => {
+  it("shows an estimated tok/s while prompting", () => {
+    fakeNow = 1000
+    vi.spyOn(performance, "now").mockImplementation(() => fakeNow)
+    mock.setConn(conn([{ type: "text", text: "a".repeat(80) }]))
+    renderBadge()
+    expect(screen.queryByText(/\d+\.\d tok\/s/)).toBeNull()
+
+    fakeNow = 2000
+    mock.setConn(conn([{ type: "text", text: "a".repeat(480) }]))
+    act(() => mock.emit())
+    expect(screen.getByText(/100\.0 tok\/s/)).toBeTruthy()
+  })
+
+  it("counts thinking blocks and skips sub-agent blocks", () => {
+    fakeNow = 1000
+    vi.spyOn(performance, "now").mockImplementation(() => fakeNow)
+    mock.setConn(
+      conn([
+        { type: "text", text: "a".repeat(80) },
+        { type: "thinking", text: "a".repeat(80) },
+        { type: "text", text: "a".repeat(800), parentToolUseId: "pt-1" },
+      ])
+    )
+    renderBadge()
+
+    fakeNow = 2000
+    mock.setConn(
+      conn([
+        { type: "text", text: "a".repeat(160) },
+        { type: "thinking", text: "a".repeat(160) },
+        { type: "text", text: "a".repeat(900), parentToolUseId: "pt-1" },
+      ])
+    )
+    act(() => mock.emit())
+    // 160 visible root chars → 40 tokens in the second second (plus the
+    // first second's 40 seeded as baseline): expect 40.0 tok/s.
+    expect(screen.getByText(/40\.0 tok\/s/)).toBeTruthy()
+  })
+
+  it("hides when the connection is not prompting", () => {
+    const idle = conn([])
+    idle.status = "connected"
+    mock.setConn(idle)
+    renderBadge()
+    expect(screen.queryByText(/\d+\.\d tok\/s/)).toBeNull()
+  })
+
+  it("hides when there is no live message", () => {
+    const noLive = conn([])
+    noLive.liveMessage = null
+    mock.setConn(noLive)
+    renderBadge()
+    expect(screen.queryByText(/\d+\.\d tok\/s/)).toBeNull()
+  })
+})
diff --git a/src/components/chat/token-output-speed.tsx b/src/components/chat/token-output-speed.tsx
new file mode 100644
index 000000000..8e6eb7624
--- /dev/null
+++ b/src/components/chat/token-output-speed.tsx
@@ -0,0 +1,97 @@
+"use client"
+
+import { useEffect, useRef, useState } from "react"
+import { Zap } from "lucide-react"
+import { useTranslations } from "next-intl"
+import { useConnectionStore } from "@/contexts/acp-connections-context"
+import { estimateTokens, TokenSpeedTracker } from "@/lib/token-speed"
+
+const THROTTLE_MS = 500
+const MAX_TPS = 999.9
+
+/**
+ * Live token-output-speed badge for one conversation tab, rendered in the
+ * status row below the composer (before the context-window circle). Shows only
+ * while the tab's agent is prompting; the reading is a local estimate from the
+ * streamed `text` + `thinking` blocks (root agent only, no sub-agent content).
+ * ponytail: root-agent only; give sub-agent cards their own badge if cross-
+ * agent delegation speed becomes something users ask about.
+ */
+export function TokenOutputSpeed({ tabId }: { tabId: string | null }) {
+  const t = useTranslations("Folder.statusBar.tokens")
+  const store = useConnectionStore()
+  const [tps, setTps] = useState<number | null>(null)
+  const lastRenderRef = useRef(0)
+
+  useEffect(() => {
+    if (!tabId) return
+    const tracker = new TokenSpeedTracker()
+    lastRenderRef.current = 0
+
+    // Mid-turn attach (refresh / reconnect): seed the baseline from whatever
+    // has already streamed so the first delta after mount measures a real
+    // rate instead of the whole accumulated text.
+    const conn = store.getConnection(tabId)
+    const content =
+      conn?.status === "prompting" && conn.liveMessage
+        ? conn.liveMessage.content
+        : null
+    if (content && content.length > 0) {
+      let total = 0
+      for (const block of content) {
+        if (
+          (block.type === "text" || block.type === "thinking") &&
+          block.parentToolUseId == null
+        ) {
+          total += estimateTokens(block.text)
+        }
+      }
+      tracker.observe(total, performance.now())
+    }
+
+    return store.subscribeKey(tabId, () => {
+      const next = store.getConnection(tabId)
+      const nextContent =
+        next?.status === "prompting" && next.liveMessage
+          ? next.liveMessage.content
+          : null
+      // Turn ended, or a fresh turn started with empty content (STATUS_CHANGED
+      // to prompting rebuilds liveMessage) — reset the baseline and hide.
+      if (!nextContent || nextContent.length === 0) {
+        tracker.reset()
+        lastRenderRef.current = 0
+        setTps(null)
+        return
+      }
+      let total = 0
+      for (const block of nextContent) {
+        if (
+          (block.type === "text" || block.type === "thinking") &&
+          block.parentToolUseId == null
+        ) {
+          total += estimateTokens(block.text)
+        }
+      }
+      const now = performance.now()
+      const rate = tracker.observe(total, now)
+      if (rate == null) return
+      if (now - lastRenderRef.current >= THROTTLE_MS) {
+        lastRenderRef.current = now
+        setTps(Math.min(Math.max(rate, 0), MAX_TPS))
+      }
+    })
+  }, [store, tabId])
+
+  if (tps == null) return null
+
+  return (
+    <span
+      aria-label={t("outputSpeedAria")}
+      className="flex items-center gap-1"
+      title={t("outputSpeedTooltip")}
+    >
+      <Zap className="size-3.5" />
+      {tps.toFixed(1)} tok/s
+    </span>
+  )
+}
diff --git a/src/i18n/messages/ar.json b/src/i18n/messages/ar.json
index e3b720214..aa85b3b5a 100644
--- a/src/i18n/messages/ar.json
+++ b/src/i18n/messages/ar.json
@@ -1713,6 +1713,8 @@
       "tokens": {
         "contextWindowUsageAria": "استخدام نافذة السياق",
         "contextWindow": "نافذة السياق",
+        "outputSpeedAria": "السرعة التقديرية للإخراج",
+        "outputSpeedTooltip": "السرعة التقديرية للإخراج (نص + تفكير)",
         "usedMax": "المستخدم / الحد الأقصى",
         "tokenUsage": "استخدام الرموز",
         "input": "إدخال",
diff --git a/src/i18n/messages/de.json b/src/i18n/messages/de.json
index b63fe4f12..dc4456463 100644
--- a/src/i18n/messages/de.json
+++ b/src/i18n/messages/de.json
@@ -1713,6 +1713,8 @@
       "tokens": {
         "contextWindowUsageAria": "Kontextfenster-Nutzung",
         "contextWindow": "Kontextfenster",
+        "outputSpeedAria": "Geschätzte Ausgabegeschwindigkeit",
+        "outputSpeedTooltip": "Geschätzte Ausgabegeschwindigkeit (Text + Denken)",
         "usedMax": "Verwendet / Max",
         "tokenUsage": "Token-Nutzung",
         "input": "Eingabe",
diff --git a/src/i18n/messages/en.json b/src/i18n/messages/en.json
index 2027ac0eb..4bd494fbc 100644
--- a/src/i18n/messages/en.json
+++ b/src/i18n/messages/en.json
@@ -1713,6 +1713,8 @@
       "tokens": {
         "contextWindowUsageAria": "Context window usage",
         "contextWindow": "Context Window",
+        "outputSpeedAria": "Estimated output speed",
+        "outputSpeedTooltip": "Estimated output speed (text + thinking)",
         "usedMax": "Used / Max",
         "tokenUsage": "Token Usage",
         "input": "Input",
diff --git a/src/i18n/messages/es.json b/src/i18n/messages/es.json
index 126886136..8ce7c5c27 100644
--- a/src/i18n/messages/es.json
+++ b/src/i18n/messages/es.json
@@ -1713,6 +1713,8 @@
       "tokens": {
         "contextWindowUsageAria": "Uso de la ventana de contexto",
         "contextWindow": "Ventana de contexto",
+        "outputSpeedAria": "Velocidad de salida estimada",
+        "outputSpeedTooltip": "Velocidad de salida estimada (texto + razonamiento)",
         "usedMax": "Usado / Máx",
         "tokenUsage": "Uso de tokens",
         "input": "Entrada",
diff --git a/src/i18n/messages/fr.json b/src/i18n/messages/fr.json
index 4fdac37d1..786f958fa 100644
--- a/src/i18n/messages/fr.json
+++ b/src/i18n/messages/fr.json
@@ -1713,6 +1713,8 @@
       "tokens": {
         "contextWindowUsageAria": "Utilisation de la fenêtre de contexte",
         "contextWindow": "Fenêtre de contexte",
+        "outputSpeedAria": "Vitesse de sortie estimée",
+        "outputSpeedTooltip": "Vitesse de sortie estimée (texte + raisonnement)",
         "usedMax": "Utilisé / Max",
         "tokenUsage": "Utilisation des tokens",
         "input": "Entrée",
diff --git a/src/i18n/messages/ja.json b/src/i18n/messages/ja.json
index 2e13d588a..73ef43d72 100644
--- a/src/i18n/messages/ja.json
+++ b/src/i18n/messages/ja.json
@@ -1713,6 +1713,8 @@
       "tokens": {
         "contextWindowUsageAria": "コンテキストウィンドウ使用率",
         "contextWindow": "コンテキストウィンドウ",
+        "outputSpeedAria": "推定出力速度",
+        "outputSpeedTooltip": "推定出力速度（テキスト + 思考）",
         "usedMax": "使用 / 最大",
         "tokenUsage": "トークン使用量",
         "input": "入力",
diff --git a/src/i18n/messages/ko.json b/src/i18n/messages/ko.json
index 95d050c9d..49b27feca 100644
--- a/src/i18n/messages/ko.json
+++ b/src/i18n/messages/ko.json
@@ -1713,6 +1713,8 @@
       "tokens": {
         "contextWindowUsageAria": "컨텍스트 윈도우 사용량",
         "contextWindow": "컨텍스트 윈도우",
+        "outputSpeedAria": "예상 출력 속도",
+        "outputSpeedTooltip": "예상 출력 속도（텍스트 + 추론）",
         "usedMax": "사용 / 최대",
         "tokenUsage": "토큰 사용량",
         "input": "입력",
diff --git a/src/i18n/messages/pt.json b/src/i18n/messages/pt.json
index 41ca4e8e9..170663563 100644
--- a/src/i18n/messages/pt.json
+++ b/src/i18n/messages/pt.json
@@ -1713,6 +1713,8 @@
       "tokens": {
         "contextWindowUsageAria": "Uso da janela de contexto",
         "contextWindow": "Janela de contexto",
+        "outputSpeedAria": "Velocidade de saída estimada",
+        "outputSpeedTooltip": "Velocidade de saída estimada (texto + raciocínio)",
         "usedMax": "Usado / Máx",
         "tokenUsage": "Uso de tokens",
         "input": "Entrada",
diff --git a/src/i18n/messages/zh-CN.json b/src/i18n/messages/zh-CN.json
index b65f15e46..844420c03 100644
--- a/src/i18n/messages/zh-CN.json
+++ b/src/i18n/messages/zh-CN.json
@@ -1713,6 +1713,8 @@
       "tokens": {
         "contextWindowUsageAria": "上下文窗口使用率",
         "contextWindow": "上下文窗口",
+        "outputSpeedAria": "估算输出速度",
+        "outputSpeedTooltip": "估算输出速度（正文 + 思考）",
         "usedMax": "已用 / 上限",
         "tokenUsage": "Token 用量",
         "input": "输入",
diff --git a/src/i18n/messages/zh-TW.json b/src/i18n/messages/zh-TW.json
index 57af307c9..fa8ec5da4 100644
--- a/src/i18n/messages/zh-TW.json
+++ b/src/i18n/messages/zh-TW.json
@@ -1713,6 +1713,8 @@
       "tokens": {
         "contextWindowUsageAria": "上下文視窗使用率",
         "contextWindow": "上下文視窗",
+        "outputSpeedAria": "估算輸出速度",
+        "outputSpeedTooltip": "估算輸出速度（正文 + 思考）",
         "usedMax": "已用 / 上限",
         "tokenUsage": "Token 用量",
         "input": "輸入",
diff --git a/src/lib/token-speed.test.ts b/src/lib/token-speed.test.ts
new file mode 100644
index 000000000..bbf9ea4a7
--- /dev/null
+++ b/src/lib/token-speed.test.ts
@@ -0,0 +1,73 @@
+import { describe, expect, it } from "vitest"
+import { estimateTokens, TokenSpeedTracker } from "./token-speed"
+
+describe("estimateTokens", () => {
+  it("counts latin at 4 chars per token", () => {
+    expect(estimateTokens("hello world")).toBeCloseTo(10 / 4)
+  })
+
+  it("counts cjk at 1.8 chars per token", () => {
+    expect(estimateTokens("你好世界")).toBeCloseTo(4 / 1.8)
+  })
+
+  it("skips whitespace", () => {
+    expect(estimateTokens("  a b  ")).toBeCloseTo(2 / 4)
+  })
+
+  it("mixes cjk and latin", () => {
+    expect(estimateTokens("你好 world")).toBeCloseTo(2 / 1.8 + 5 / 4)
+  })
+
+  it("counts fullwidth punctuation as cjk", () => {
+    expect(estimateTokens("你好，世界！")).toBeCloseTo(6 / 1.8)
+  })
+
+  it("returns 0 for empty or whitespace-only text", () => {
+    expect(estimateTokens("")).toBe(0)
+    expect(estimateTokens("   ")).toBe(0)
+  })
+})
+
+describe("TokenSpeedTracker", () => {
+  it("seeds the baseline on the first observation", () => {
+    const tracker = new TokenSpeedTracker()
+    expect(tracker.observe(0, 0)).toBeNull()
+  })
+
+  it("computes a constant rate", () => {
+    const tracker = new TokenSpeedTracker()
+    tracker.observe(0, 0)
+    expect(tracker.observe(100, 1000)).toBeCloseTo(100)
+  })
+
+  it("smooths a burst followed by silence", () => {
+    const tracker = new TokenSpeedTracker()
+    tracker.observe(0, 0)
+    tracker.observe(200, 1000)
+    const rate = tracker.observe(200, 2000)
+    expect(rate).toBeGreaterThan(0)
+    expect(rate).toBeLessThan(200)
+  })
+
+  it("decays toward zero over a long pause", () => {
+    const tracker = new TokenSpeedTracker()
+    tracker.observe(0, 0)
+    tracker.observe(100, 1000)
+    expect(tracker.observe(100, 10_000)).toBeLessThan(1)
+  })
+
+  it("ignores non-positive time deltas", () => {
+    const tracker = new TokenSpeedTracker()
+    tracker.observe(0, 1000)
+    expect(tracker.observe(50, 1000)).toBeNull()
+    expect(tracker.observe(100, 500)).toBeNull()
+  })
+
+  it("resets to a fresh baseline", () => {
+    const tracker = new TokenSpeedTracker()
+    tracker.observe(0, 0)
+    tracker.observe(100, 1000)
+    tracker.reset()
+    expect(tracker.observe(0, 2000)).toBeNull()
+  })
+})
diff --git a/src/lib/token-speed.ts b/src/lib/token-speed.ts
new file mode 100644
index 000000000..cb460ed6a
--- /dev/null
+++ b/src/lib/token-speed.ts
@@ -0,0 +1,59 @@
+/**
+ * Pure helpers behind the live token-output-speed badge. Everything here is a
+ * function of its arguments — no clock, no DOM — so the heuristic and the
+ * smoothing are exactly testable.
+ *
+ * The char→token ratios are deliberately rough ("大差不差"): CJK text
+ * tokenizes denser than Latin, so we count CJK (ideographs + fullwidth forms)
+ * at ~1.8 chars/token and every other visible character at ~4 chars/token.
+ * Whitespace is skipped entirely.
+ * ponytail: fixed ratios; calibrate per model from reported turn usage if
+ * accuracy ever matters more than the live gauge.
+ */
+
+const CJK_RE = /[\u3400-\u4DBF\u4E00-\u9FFF\uF900-\uFAFF\uFF00-\uFFEF]/g
+
+export function estimateTokens(text: string): number {
+  if (!text) return 0
+  const cjk = text.match(CJK_RE)?.length ?? 0
+  const other = text.replace(/\s/g, "").length - cjk
+  return cjk / 1.8 + other / 4
+}
+
+/**
+ * First-order low-pass over instantaneous token rates. Time-constant based
+ * (rather than per-event alpha) so a bursty thinking stream followed by a tool
+ * pause decays toward zero instead of pinning the old reading.
+ */
+export class TokenSpeedTracker {
+  private static readonly TAU_MS = 1500
+
+  private lastTime: number | null = null
+  private lastTokens = 0
+  private ewma: number | null = null
+
+  reset(): void {
+    this.lastTime = null
+    this.lastTokens = 0
+    this.ewma = null
+  }
+
+  /** Feed the cumulative estimated token count at `nowMs`; returns smoothed
+   *  tok/s, or `null` while seeding / when the observation can't be used. */
+  observe(totalTokens: number, nowMs: number): number | null {
+    if (this.lastTime == null) {
+      this.lastTime = nowMs
+      this.lastTokens = totalTokens
+      return null
+    }
+    const dt = nowMs - this.lastTime
+    if (dt <= 0) return this.ewma
+    const instant = (totalTokens - this.lastTokens) / (dt / 1000)
+    this.lastTime = nowMs
+    this.lastTokens = totalTokens
+    const alpha = 1 - Math.exp(-dt / TokenSpeedTracker.TAU_MS)
+    this.ewma =
+      this.ewma == null ? instant : this.ewma * (1 - alpha) + instant * alpha
+    return this.ewma
+  }
+}