OpenAI has added full-duplex voice to Codex, which means you can now debate a pull request, sketch an architecture, check progress, or change direction mid-task without dropping back to the keyboard.
You keep talking. Codex keeps working.
That is the real shift.
Codex was already an agent capable of writing, reviewing, debugging, and working across threads. Voice turns it into something closer to a control layer you can speak to.
Instead of constantly switching between thinking, typing, checking, and redirecting, you can stay in the conversation while the agent handles the execution.
ChatGPT Voice, powered by GPT-Live, now works with Codex and ChatGPT Work in the desktop app, letting you hold a live spoken conversation while starting, checking, and steering coding agents in the background.
This is very different from the old “hold spacebar and dictate a prompt” experience in the Codex CLI.
Dictation converts speech into text.
Voice is an ongoing conversation that can actually move work forward.
GPT-Live is a full-duplex speech model. It can listen while speaking, handle interruptions, and maintain a natural back-and-forth instead of forcing rigid turn-taking.
That matters because the voice model does not need to perform every heavy coding task itself.
Long-running reasoning, debugging, testing, and implementation can be handed off to Codex tasks and background models while the voice layer stays responsive.
You get the speed of a conversation without forcing the actual coding work to happen at conversational speed.
Guinness Chen, who works on Codex and ChatGPT Voice at OpenAI, has described a workflow that captures the idea well:
The longer-term vision is even more interesting: stop manually managing threads and agents.
Instead, describe the outcome you want and let Voice coordinate the work required to get there.
Inside the ChatGPT desktop app, you can open an existing Codex task and start talking, or begin a new voice conversation and dispatch work from there.
You might say:
Voice can launch separate tasks for longer jobs, send follow-up instructions, and bring status updates, blockers, and results back into the same conversation.
You no longer need to babysit a terminal just to stay informed.
On macOS, Voice can also use screen context.
Enable it under Settings → Voice, then say something like:
“Take a look at this.”
ChatGPT can capture an appshot of the frontmost window, including the image and accessible text. In some cases, that context can include content outside the currently visible scroll area.
That makes it much easier to discuss code, errors, dashboards, designs, or documents without manually copying everything into the conversation.
It also means screen context should be treated as a permission surface.
Be careful what is visible when you invoke it, especially around credentials, customer data, private messages, or sensitive internal tools.
The desktop rollout began July 23, 2026 for Plus, Pro, Business, Edu, and Enterprise users, with Edu and Enterprise receiving a short early-access window.
Workspace administrators can still disable or hide some functionality.
The mobile Remote experience is where this starts to feel less like a demo and more like a genuine workflow.
Pair your Mac or Windows machine, open ChatGPT on your phone, go to Remote, and start realtime voice control.
Now you can steer Codex while walking, commuting, or moving between meetings.
The files, credentials, repository, and development environment stay on the host machine. You simply send instructions and receive updates through your phone.
That is a much more compelling version of “hands-free coding” than speaking prompts into a laptop while sitting at the same desk.
The simplest mental model is:
Use Voice for planning and steering. Use Codex for execution.
Talk through the problem until the idea is clear enough to build.
Ask questions. Challenge assumptions. Refine the architecture. Decide what “done” actually means.
Then launch a dedicated Codex task on a stronger, longer-thinking model for the implementation.
Keep the voice thread focused on direction:
“Status?”
“Why did that test fail?”
“Try the other approach.”
“Look at this window.”
“Open a PR when everything is green.”
That division is intentional.
Chen has said the live Voice experience uses a lower-latency backing model so conversations remain responsive. Putting the largest reasoning model directly inside the realtime audio loop would make speaking with it feel sluggish.
So the harder work gets pushed into dedicated tasks.
If you only need to dump a prompt, use dictation.
If you need to think out loud while moving work forward, use Voice.
There are still practical constraints.
None of these undermine the feature.
They simply make the difference between “Jarvis in a marketing demo” and an actual product with permissions, quotas, and infrastructure.
Coding agents already changed the unit of software work.
The old model was:
Write this function.
The newer model became:
Own this task. Investigate it. Fix it. Report back.
Voice changes the unit of control.
You can now stay in a design discussion, walk through an idea, sit in a meeting, or step away from your desk while continuing to redirect an agent that is actively working inside the repository.
For developers already using Codex, the quality-of-life improvement is obvious: less typing, fewer status-check loops, less context switching, and a shorter distance between a fuzzy idea and an executing task.
But the more interesting change is cultural.
Software development starts to look less like operating tools one command at a time and more like directing a small team.
You explain what should happen.
The agents go work.
They return with questions, blockers, and results.
You keep the conversation moving at human speed.
That model only works if the agent is competent enough that directing it is cheaper than doing the work yourself.
For review, scaffolding, debugging, and increasingly large chunks of implementation, Codex is already approaching that threshold.
Voice removes another layer of friction between:
“I know what I want.”
and
“Go make it happen.”
Official reference: ChatGPT Voice in Codex
https://developers.openai.com/codex/features/voice
The interesting part is not that an AI can talk.
Plenty of products can talk.
The interesting part is that the thing you are talking to can still be shipping code while you finish the sentence.
| Surface | What you get |
|---|
| ChatGPT desktop on macOS and Windows | Native Voice across Chat, Work, and Codex |
| Existing Codex threads | Voice support as the rollout reaches updated desktop apps and task hosts |
| iOS via Remote | Pair your phone with a desktop host and talk to Codex through ChatGPT |
| Android via Remote | Similar remote workflow, following the iOS rollout |
| Web or standalone mobile Voice | Not the same agent-control surface; the real work runs through desktop or a remote desktop host |