diff --git a/src/voice/gevRealtime.test.mjs b/src/voice/gevRealtime.test.mjs
index b1e773e49..541e2cc04 100644
--- a/src/voice/gevRealtime.test.mjs
+++ b/src/voice/gevRealtime.test.mjs
@@ -3208,6 +3208,7 @@ test('OAuth cloud sessions report unknown cost and usage without tripping the AP
     responses: 1,
     input: 0,
     output: 1562500,
+    captions: 0,
     costKnown: false,
   });
   const after = controller.costTracker.state();
@@ -3220,6 +3221,28 @@ test('OAuth cloud sessions report unknown cost and usage without tripping the AP
   assert.match(ui.costValue.title, /Usage is incomplete/);
 });
 
+test('OAuth cloud sessions count caption transcriptions without pricing them into the API cap', () => {
+  const { controller, ui } = costControllerHarness();
+  controller.status = 'listening';
+  controller._cost.cloudVoiceAuth = 'oauth';
+  controller._cost.sessionCloudVoiceAuth = 'oauth';
+  controller._cost.bindTranscriptionModel('gpt-4o-mini-transcribe');
+  const before = controller.costTracker.state();
+  // Priced, this much caption audio would cross the $5 cap.
+  const result = controller.recordTranscriptionUsage({
+    type: 'duration',
+    seconds: 3_600_000,
+  });
+  assert.equal(result.captions, 1);
+  assert.equal(result.costKnown, false);
+  const after = controller.costTracker.state();
+  assert.equal(after.totalUsd, before.totalUsd);
+  assert.equal(after.capReached, false);
+  assert.equal(controller.costCapStopped, false);
+  assert.equal(ui.costValue.textContent, 'COST UNKNOWN');
+  assert.match(ui.costValue.title, /1 caption transcription\(s\)/);
+});
+
 test('F1: when idle, toggling does re-price the preview meter', () => {
   const { controller } = costControllerHarness();
   controller.status = 'idle';
diff --git a/src/voice/realtimeCost.js b/src/voice/realtimeCost.js
index 0d48a7701..d3d7f684e 100644
--- a/src/voice/realtimeCost.js
+++ b/src/voice/realtimeCost.js
@@ -20,7 +20,7 @@ export class RealtimeCost {
     this.voiceLimits = readStoredVoiceLimits();
     this.cloudVoiceAuth = readStoredCloudVoiceAuthMode();
     this.sessionCloudVoiceAuth = null;
-    this.oauthUsage = { responses: 0, input: 0, output: 0 };
+    this.oauthUsage = { responses: 0, input: 0, output: 0, captions: 0 };
     this.transcribeModel = null;
     this.costTracker = createVoiceCostTracker({
       tier: this.voiceTier,
@@ -64,12 +64,12 @@ export class RealtimeCost {
     if (this.ui?.costValue) {
       const liveAuth = this.sessionCloudVoiceAuth || this.cloudVoiceAuth;
       if (liveAuth === 'oauth') {
-        const { responses, input, output } = this.oauthUsage;
+        const { responses, input, output, captions } = this.oauthUsage;
         this.ui.costValue.hidden = false;
         this.ui.costValue.textContent = 'COST UNKNOWN';
         this.ui.costValue.dataset.level = 'unknown';
         this.ui.costValue.title =
-          `ChatGPT OAuth session: ${responses} response(s), ${input} input and ${output} output tokens reported. ` +
+          `ChatGPT OAuth session: ${responses} response(s), ${input} input and ${output} output tokens, ${captions} caption transcription(s) reported. ` +
           'USD cost and the API spend cap are unavailable for this auth mode.' +
           (state.incomplete
             ? ' Usage is incomplete because a response was still in flight when the session ended.'
@@ -176,6 +176,13 @@ export class RealtimeCost {
    */
   recordTranscriptionUsage(usage) {
     if (!usage) return null;
+    // An OAuth session has no API-dollar meter, so captions are counted
+    // rather than priced into the API-key warning and cap.
+    if (this.sessionCloudVoiceAuth === 'oauth') {
+      this.oauthUsage.captions += 1;
+      this.syncCostUi();
+      return { ...this.oauthUsage, costKnown: false };
+    }
     return this.applyCost(
       this.costTracker.recordUsd(
         estimateTranscriptionCostUsd(usage, this.transcribeModel),
@@ -243,7 +250,7 @@ export class RealtimeCost {
     this.voiceLimits = readStoredVoiceLimits();
     this.cloudVoiceAuth = readStoredCloudVoiceAuthMode();
     this.sessionCloudVoiceAuth = this.cloudVoiceAuth;
-    this.oauthUsage = { responses: 0, input: 0, output: 0 };
+    this.oauthUsage = { responses: 0, input: 0, output: 0, captions: 0 };
     this.costCapStopped = false;
     // Provisional meter (tier-priced) so the readout shows $0.00 while
     // connecting. It is REPLACED below with one bound to the model the server