Quick Facts
- OpenAI integrated GPT-Live’s full-duplex voice into its desktop app on July 23, 2026, in build 26.715 for macOS and Windows.
- Developers can now speak a single prompt to trigger multiple simultaneous Codex tasks, including bug investigation, pull request review, and test generation.
- The feature is available to Plus, Pro, Business, Edu, and Enterprise subscribers and reaches more than 10 million weekly active users across Codex and ChatGPT Work.
OpenAI brought hands-free voice control to its unified desktop app on July 23, 2026, wiring GPT-Live’s full-duplex audio model directly into Codex and ChatGPT Work. The update, shipped in desktop build 26.715, lets developers issue spoken commands that trigger multiple concurrent agent tasks without touching a keyboard.
The release came two weeks after OpenAI first unveiled GPT-Live on July 8, a continuous audio model that listens and speaks at the same time. That eliminated the rigid turn-taking of earlier voice systems. The July 23 update brings that capability into the same desktop surface that merged Chat, Work, and Codex on July 9, retiring the standalone Codex app.
How the System Works
Full-duplex means both the user and the model can transmit at the same time. GPT-Live manages the live audio layer, inserting verbal acknowledgments while passing heavy reasoning work to GPT-5.6 Terra, a newer background model than the GPT-5.5 pairing from the initial July 8 launch.
On macOS, the app uses an “Appshots” feature and screen context to analyze the frontmost window alongside local files, codebase structures, and active plugins. That setup allows ChatGPT Voice to track what a developer is looking at and act on it without a manual handoff.
Build 26.715 also shipped multi-folder local projects. A primary folder drives new chats, Git operations, and automatic discovery of configuration files, while secondary folders remain accessible for file reads.
What Developers Can Do
A single spoken prompt can now start multiple task threads at once. OpenAI’s own demo showed a developer instructing the system to investigate an authentication bug, review an API migration pull request, and generate missing unit tests in one command. The desktop app coordinates those actions across Slack conversations, GitHub repositories, and local codebases.
A separate demo cited by 9to5Mac showed one voice command create a task thread, open a pull request, and identify a bug’s root cause. Developers can also convert design mockups into working code verbally, splitting work across frontend, backend, and testing layers.
OpenAI defined two distinct modes in its release notes. “Voice in Chat” covers real-time conversation, brainstorming, and web search across desktop, iOS, and Android. “Voice in Work and Codex” starts on the desktop app and lets ChatGPT Voice control the computer and coordinate work across multiple agents, active conversations, and projects. Remote access via the ChatGPT iOS app is supported now; Android support is coming.
Pricing and Access
The feature is available to Plus, Pro, Business, Edu, and Enterprise subscribers. Free and Go tiers are excluded. Voice minutes replenish on a rolling five-hour cycle, separate from the Codex task budget. Business, Edu, and Enterprise workspaces on credit or pay-as-you-go billing pay roughly 6 credits per minute of desktop Voice. Tasks started through voice consume standard Codex and ChatGPT Work quotas, the same as any other agentic workload.
OpenAI Codex engineer Tibo Sottiaux previewed the release the day before on X, writing: “Unbelievable excited for what’s coming together. Tomorrow is feeling codexy.” An early user described the experience: “It listens while it talks, and it keeps working while you talk. So you kick off one job, start describing the next one out loud, and it stays with you instead of freezing until the first one finishes.”
The update reaches more than 10 million weekly active users across Codex and ChatGPT Work, the largest distribution of agentic voice controls OpenAI has shipped to date.
This article was written by an AI agent. Spotted an error? Send a correction and we will fix it.