On September 23, 2026, OpenAI expanded ChatGPT's voice features so that plugins can be used from Live, its real-time conversation mode. The update covers web, iOS and Android. In addition, ChatGPT Work, the feature for drafting documents and carrying out research, can now be operated by voice on web and mobile. Voice requests for work, which previously started on the desktop, have now reached smartphones. However, being able to give an instruction by voice is not the same as being able to carry it out in a connected service. You need to look separately at which apps are available, what permissions apply, and how approvals are handled.
Calling connected apps and handing tasks to Work
With this update, you can call plugins and connected apps available to your account during a Live conversation. OpenAI's release notes explain that you can talk while also following the responses as text in the chat. Information you get mid-conversation can be checked on screen rather than only being heard again.
The other change is voice support for ChatGPT Work. Work is a feature for taking on longer jobs than a one-off question, using connected apps and files to carry a task from research through to a finished deliverable. From September 23, you can hand tasks to Work by voice on web and mobile as well. The tasks OpenAI lists include creating documents, presentations and spreadsheets, as well as work that uses connected apps or a browser.
One conceivable use is asking for a document while on the move and refining its contents by speaking. If the task isn't finished when the call ends, OpenAI says you can continue in a text conversation. That said, this article did not test a full workflow with any specific plugin on an actual device.
Voice control that began on desktop now reaches web and phones
The ability to run Work by voice is not itself new. On July 23, OpenAI introduced voice to Work and Codex in the desktop app, letting users start tasks and check progress by speaking. Work itself was announced on July 9 as a feature for research and document creation. What is new this time is that voice control of Work has been extended to web and mobile, and that plugins can now be connected to Live in regular Chat.
Knowing this background makes the feature's position clearer. Plugins are a mechanism for using connections to external services and set procedures in ChatGPT, while Work is where multi-step jobs get done. The September 23 update adds more ways to reach both by voice. It is not an announcement that grants new access to email or calendars you haven't connected, nor that opens up every app's actions across the board.
Model names also call for care. OpenAI's Voice help page says the model behind Live conversations is GPT-Live-1 or GPT-Live-1 mini, depending on the plan. It also explains that other models can be used for difficult searches and reasoning, and on September 22 OpenAI introduced GPT-6 Sol and Luna to Work and Codex. Lumping together the models selectable in Work and the model that runs real-time voice, and reading it as "voice runs on GPT-6," overlooks the difference in their roles.
You can ask by voice, but approvals stay on screen
To use a plugin in Live, you may need to connect the app and sign in. The information and actions available inherit the permissions of the connected account and any workspace restrictions. Speaking does not widen those boundaries. Which reads and writes are possible in a connected service such as email, calendar or Slack depends on the actions and settings each app provides.
Actions that require approval show a confirmation on the web or mobile screen. OpenAI says voice approval is not supported and directs users to allow or deny on screen. In other words, you can start a task by speaking, but work that involves sending something externally or making changes won't always finish without looking at the screen. What you have permitted a connected service to do, and which actions you will be asked to confirm before they run, are the practical points that matter.
Plan differences are also significant. Free and Go users can use supported plugins from Voice within Chat on their plans. Voice within Work, on the other hand, requires access to both Voice and Work. Time spent using Work by voice follows the plan's Voice usage limits and billing, and Work tasks that are run count toward normal Work usage. Requesting by voice does not mean the Work allowance becomes unnecessary. Available features also vary by region, app version and organization settings, so check whether Work and the relevant plugins appear on your own screen.
Before you start, it helps to sort what you want done into "finding information," "creating a deliverable" and "changing the state of a connected service." Even for the first two, results won't match expectations without the necessary materials and access to connected services. For the last, check whether the service offers write actions, whether your account has the permissions, and whether an on-screen approval will be requested. A list of features you can request by voice does not determine the range of work that can actually be completed.
If you handle personal schedules or email by voice, it is also worth reviewing how conversation data is managed. According to OpenAI's Voice help, audio clips from Live and Advanced are stored alongside the transcripts shown in chat history and kept for 30 days. Audio clips are not used to train models unless the user chooses to share them. Transcripts and files shared during a conversation, on the other hand, may be used for training depending on the plan and data settings. Before adding more connected apps, check the data settings for your account and organization.
Your choice of voice mode changes what you can do
ChatGPT's "voice" is no longer a single thing. The table below arranges the mode descriptions from the official help page by the situations in which you would use them.
| Voice mode | How the conversation works and connected apps | Video / screen sharing |
|---|---|---|
| Live | Converses in real time and can use available plugins and connected apps | Not supported |
| Advanced | The previous real-time voice. The official help page does not state plugin support | Available on supported mobile versions |
| Standard | Turn-based: converts your speech to text before responding. The official help page does not state plugin support | Not stated in the official help page |
Live can use available plugins by voice but does not support video or screen sharing. Advanced can use video and screen sharing on supported mobile devices. Voice in Work requires access to both features, plus existing permissions.
The "both features" here refers to Voice and Work, which OpenAI's release notes list as the conditions for use.
This table lines up the features OpenAI describes; it is not a comparison that measures each mode's capabilities under the same conditions. "Not stated" does not mean unsupported, nor does it mean every mode can be selected on every plan. Whether you want to use connected apps by voice or show your camera or screen determines which mode to use and which conditions to check.
The usefulness of ChatGPT's voice can no longer be judged just by how long you can talk. What remains to be seen is which actions are actually permitted in the apps users connect, and whether they can move smoothly to their smartphone screen when approval is needed. Only after checking that can you judge how much of a task started by voice can be entrusted to it.