On September 23, OpenAI added plugin support to Live on web, iOS and Android. Voice conversations can now work with available connected apps.
For users, this moves a conversation about a task closer to working with the relevant material. You might ask a question about a document aloud, then check the written answer. Its value depends on whether the required app is available and whether reviewing the result takes less time than an ordinary search.
What changes in Chat and Work
According to the ChatGPT Voice guide, voice in Work can start tasks involving documents, presentations, spreadsheets and a browser; unfinished work can continue in text after the call ends. Both Voice and Work access are required, and normal usage limits still apply.
Treat voice as another way to give a specific task. It can be useful when explaining an idea aloud is easier, while the final document, calculation or list still needs review.
Installing a plugin and authorizing an account are separate steps
A plugin can combine instructions and apps. An app connects an external service and works within its capabilities and granted permissions. Installing a plugin does not replace authorization for the intended account or override workspace restrictions.
To connect, open Plugins or Apps as available in your interface, review the included capabilities, install the plugin if needed, and complete the requested connection. The specific capability depends on plan, region, role and the surface you use.
Start with one service you already use. Before connecting it, identify the file or kind of information needed for a first task. This makes the integration’s usefulness and requested access easier to assess.
Action approval still happens on screen
When an action requires approval, Live shows an on-screen request; spoken approval is unsupported. Live also does not support video or screen sharing. The guide separately identifies Advanced Voice for those supported mobile capabilities.
Before approving, check the item being changed, the account, recipient and final text, particularly if a filename or name was spoken unclearly. A search or a draft you can read at your own pace makes a sensible first attempt.
Two examples to try first
Working with a document. “In my connected storage, find the Launch Plan file. List the three nearest deadlines and their owners, and link to the source. If a date is missing, say so.” The filename is fictional; replace it with your own.
Preparing material in Work. “Turn these notes into a five-slide presentation draft. Separate facts from assumptions and flag unanswered questions.” These are suggested requests, not results from our own testing of the feature.
- Check that the correct file and its current version were found.
- Compare numbers, names and deadlines with the original.
- Measure time to a usable result including review, rather than only the speed of the spoken answer.
What happens to the data
The connected-apps guide says data shared with an app follows that service’s rules. For personal Free, Go, Plus and Pro accounts, app information may be used for model training when Improve the model for everyone is enabled. Business, Enterprise and Edu app data is not used for training by default.
Disconnecting an app account stops future access through it but does not automatically erase existing conversations, saved files or memories. If an app is unavailable in Voice, try a supported text conversation; voice capabilities differ between integrations.
Before using sensitive material, review data settings and your workspace’s rules. For another approach to conversational AI, see our Gemini 3.8 Live explainer. Other developments appear in the September 21–27 digest.
Join the conversation
Stay on topic and respect other readers. Your first comment may appear after editorial review.