[macOS 26.5.2][26.721.41059] Codex dictation consistently fails with "Unable to transcribe audio" while ChatGPT Classic works

Open 💬 0 comments Opened Jul 27, 2026 by u0400587

Summary

Voice dictation in the current ChatGPT/Codex desktop app records audio but consistently fails at the transcription stage with:

Unable to transcribe audio

ChatGPT Classic dictation works on the same Mac and ChatGPT account. This isolates the failure to the new ChatGPT/Codex desktop app or its transcription path rather than the microphone.

Environment

  • Product: new ChatGPT desktop app, Codex view
  • App version: 26.721.41059
  • App build: 5848
  • Authentication: ChatGPT sign-in, not API-key auth
  • Hardware: MacBook Air M1 (arm64)
  • macOS: 26.5.2
  • Comparison app: ChatGPT Classic 1.2026.184

Steps to reproduce

  1. Open the new ChatGPT desktop app and switch to Codex.
  2. Open any task, including a new empty task.
  3. Start dictation using the microphone control or Ctrl + Shift + D.
  4. Speak for 3–5 seconds.
  5. Stop recording and wait for transcription.

Actual behavior

  • Recording starts normally.
  • No transcript is inserted.
  • The app displays Unable to transcribe audio.
  • The failure is consistently reproducible with short recordings.

Expected behavior

The recording should be transcribed and inserted into the Codex composer as editable text.

Troubleshooting performed

  • Microphone access is granted.
  • macOS system dictation works.
  • ChatGPT Classic dictation works on the same machine and account.
  • Restarting the new app and testing a new empty task did not help.
  • The installed app matches the latest production version available at the time of reporting.

Request

Please investigate the transcription path in ChatGPT/Codex desktop 26.721.41059 on Apple Silicon, including any difference from ChatGPT Classic. Please also expose the underlying transcription error code instead of only showing a generic message.

View original on GitHub ↗