OpenAI Connects Plugins to ChatGPT Voice

OpenAI significantly expanded ChatGPT Voice on September 23. Users can now call on plugins such as email, calendar, and Slack in the middle of a voice conversation, and the models doing the work behind the voice are now GPT-6 Astra, Sol, and Luna. In ChatGPT Work, the workspace for building documents, presentations, and spreadsheets, users can now hand off tasks by voice on both web and mobile. "We heard you loud and clear," OpenAI wrote in its announcement on X.

The shift is that voice has moved from a tool that answers to a channel for delegating work. According to OpenAI's release notes, the Live mode of voice conversations now supports plugins on web, iOS, and Android, and can use the apps already connected to an account during the conversation. Users can listen to the answer while following it as text in the chat window. If a task in Work is not finished when the voice call ends, it keeps running in text.

The plan decides who gets what. According to reports, Plus and Pro users can use the Work tab on their phones to create documents, draft emails, and summarize Slack messages, and can also build sites, write presentations, and use the cloud browser. Free and Go users can use only the plugins their plan supports, through voice in regular chat. Using Work by voice requires access to both Voice and Work, and existing app connections, permissions, and usage limits still apply.

According to OpenAI's help center, voice usage limits are measured over a rolling 24-hour window. Go gets 3 hours with GPT-Live-1 mini, Plus gets 3 hours with GPT-Live-1, and Pro gets 15 hours or unlimited use depending on the tier. Work tasks started by voice count toward Work usage just as typed tasks do. Users can pick from nine voices, including Arbor, Cove, and Juniper, and can also start a voice conversation from Apple CarPlay.

The safeguards live on the screen. According to the help center, when a plugin action needs the user's approval, Voice asks for confirmation on screen, and the user has to approve or decline with the on-screen buttons. Spoken approval is not accepted. Audio clips from voice conversations are kept for 30 days along with the chat history, and the company said the audio is not used for model training unless the user chooses to share it. Live mode does not yet support video or screen sharing.

A 125-second OpenAI launch video that METAL reviewed shows the change through everyday scenes. One user discovers by voice that a finance app is paying for a magazine twice, and asks in one go for the duplicate to be canceled and a refund request to be emailed. Another builds a website to sell pottery by speaking, adds a checkout page, emails a marina about a sailboat rental, and then books a boat for 4 p.m. A weather answer with winds of 8 knots and gusts up to 12 knots also comes back by voice.

The person leading the update is Atty Eleti, product lead for ChatGPT voice. According to reports, in a demo Eleti only spoke to the app to analyze his food delivery spending, book a pickleball court, and reorganize his calendar. "In the future, we really imagine people using voice as the primary modality to interact with AI, and that's kind of our North Star that we're working towards," he said. He said more than 150 million people already use dictation and voice mode.

Eleti explained that the design also accounts for the fact that voice is not always the right answer. Speaking is faster than typing, but listening is slower than reading. "We don't think of this as just pure voice in and voice out. It's about voice in and then the right output back," he said. "Sometimes it's text, sometimes it's voice, sometimes it's a mix." The setup of speaking an answer while showing text in the chat window grew out of that judgment.

The announcement extends a voice strategy that has moved in quick steps. METAL has reported that OpenAI released the GPT-Live-1 voice model, which listens and speaks at the same time, through its API, and after launching GPT-Live in July the company first connected it to the desktop app so users could run Work and Codex tasks by voice. On September 9 it let voice use GPT-5.6 or GPT-6 Astra for harder questions, and one day earlier, on September 22, it unveiled GPT-6 Sol and Luna. This update brings those models all the way into mobile voice.

The competitive pressure has also risen. According to reports, Meta's personal agent Muse passed 500,000 users and 250,000 daily active users in its first week. METAL has covered Meta's launch of Muse. OpenAI is rolling out announcements ahead of its DevDay developer event next week, and according to reports it is also building a smart speaker targeted for release in 2027. If the habit of handing off work by voice takes hold, it becomes the default way to operate that device.

From a reporter's point of view, the weight of this announcement lies in its direction more than in its feature list. OpenAI has begun putting at the center of its product a flow in which a user holding a phone sends emails, cancels charges, and finishes bookings while barely looking at the screen. At the same time, actions that need the user's approval still have to be confirmed with on-screen buttons. How long that line holds, delegating by voice but confirming by eye, is the next thing to watch.