1. Confirm you are on a surface that has dictation
OpenAI’s shortcut reference for the ChatGPT desktop app, where Codex now lives, lists Start dictation as ⌃⇧D on macOS and Ctrl+Shift+D on Windows and Linux, each marked “when dictation is available”. Click into the composer first, press the shortcut, and speak; the transcript arrives as editable text that you send yourself.
The terminal is a different product surface. As of October 2026, OpenAI’s Codex CLI references describe no dictation mode. Codex CLI 0.156.0, released September 22, 2026, turned voice conversations on by default instead, with an F8 toggle, a /voice settings picker, and bundled audio runtimes for Linux and Windows. That is a spoken back-and-forth with the agent, not speech typed into your prompt.
We also found no dictation command documented for the Codex IDE extension. If the editor or terminal is where you type, dictate there with an operating-system or third-party tool.
- Desktop app: click the composer, then press Control-Shift-D (Mac) or Ctrl+Shift+D.
- Codex CLI: press F8 for a voice conversation; use outside dictation for typed prompts.
- IDE extension: no dictation documented, so use an OS-level tool.
2. Check your plan, sign-in method and admin settings
OpenAI’s Codex feature table lists voice dictation as available on ChatGPT Plus, Pro, Business and Enterprise/Edu, and unavailable when you use an API key. The same table does not list the Free or Go plans. If you signed in to Codex with an API key, the shortcut has nothing to start.
Managed workspaces can turn it off. OpenAI’s configuration reference describes features.in_app_dictation in the admin-enforced requirements.toml: false disables in-app dictation in the desktop app, and true does not bypass other availability checks. A separate key, features.realtime_conversation, disables the experimental /voice command in the Codex CLI. If dictation is missing on a work machine but present at home, ask whoever manages your Codex configuration before looking for a workaround.
3. The “Codex dictation usage limit reached” message
OpenAI’s documentation reviewed for this page publishes no dictation quota and does not explain this message. The only first-hand record we found is a September 16, 2026 bug report in OpenAI’s developer community: macOS ChatGPT app version 26.908.70816 showed “Codex dictation usage limit reached” on a Free account, while dictation still worked in ChatGPT on the web and in the iPad app for the same account.
Two documented facts help you reason about it. Free is not a plan in the Codex dictation row, and OpenAI says live ChatGPT Voice in the desktop app draws on your Codex usage budget at $0.05 a minute and shares your plan’s usage limits. Dictation pricing itself is not documented, so treat any claim about a dictation-specific allowance as unverified, including ours.
- Compare your plan with the Codex feature table.
- Open your Codex usage dashboard for current limits and reset times; Plus and Pro can buy extra credits.
- Try dictation in ChatGPT on the web to separate an account limit from an app bug.
- Update the desktop app and report the version number if the message persists.
4. The shortcut does nothing, or no text appears
Start with the operating system. The ChatGPT app needs microphone permission: on a Mac, enable it under System Settings → Privacy & Security → Microphone; on Windows, check Settings → Privacy & security → Microphone. A microphone that works in a browser proves the hardware, not the app’s permission.
Then rule out the other voice feature. Control-Shift-V starts a live voice chat and Control-Shift-D starts dictation, so a mistyped shortcut opens the wrong mode. Shortcuts can be reassigned under Settings > Keyboard Shortcuts, so check there that Start dictation still has the binding you expect.
5. Where your dictation audio goes
OpenAI’s Voice Dictation FAQ, written for ChatGPT’s microphone dictation, says recorded audio is sent to OpenAI’s models to be transcribed and returned as text you can edit before sending. Dictation is not processed on your computer.
The same FAQ says dictation audio is kept for as long as the chat is in your history. When you delete the chat, the audio is deleted within 30 days unless OpenAI must keep it for security or legal reasons, or it was previously shared for training and disassociated from your account. Consumer users can stop sharing dictation audio for model training by turning off “Include your audio recordings” in data controls; OpenAI says it does not train on business users’ inputs by default.
6. Dictate into Codex anywhere with IraVoice
Codex’s own dictation is the right first choice in the desktop app on an eligible plan. When Codex runs in a terminal, the IDE or over SSH, when you sign in with an API key, or when dictation is capped, a dictation app on your own computer can type the prompt instead. That is the gap IraVoice fills; it is not a live voice conversation and does not replace ChatGPT Voice.
IraVoice 1.2.2 runs on macOS 13+ on Apple silicon, Windows 10 or 11 (64-bit), and Linux (beta) and recognizes speech on-device with NVIDIA Parakeet. On IraVoice Local no audio or text leaves the computer and nothing is metered per minute: it is $79 once for one computer. Hold Right Command on a Mac, Right Ctrl on Windows or Ctrl+Space on Linux, speak, and release; the text lands at the cursor and Enter is not pressed by default.
When Codex runs in a terminal or editor on a Mac or Windows PC, IraVoice also reads names from the Git project that session is working in, so “pipeline dot swift” arrives as Pipeline.swift. It keeps only names, never source text. Code names are not available on Linux, and over SSH they come only from a project on your local computer.