What changed in ChatGPT Voice on September 23, 2026?

Three things, per OpenAI's release notes. Live voice mode now supports plugins and connected apps on web, iOS and Android, so you can ask ChatGPT to read your calendar, draft an email or post to Slack mid-conversation and see the written result in the chat. Voice is now available inside ChatGPT Work on web and mobile, so you can start a document, presentation, spreadsheet or browser task by speaking and let it continue in text after the call ends. And paid users can choose which GPT-6 model, Astra, Sol or Luna, Voice escalates to when it needs to search or reason, following the September 9 change that first let Voice hand off to GPT-5.6 or GPT-6 Astra.

Which model does ChatGPT Voice use?

The real-time conversation runs on GPT-Live-1 for paid plans and GPT-Live-1 mini for Free users, both introduced July 8, 2026; they listen and speak at the same time so interruptions feel natural. When a question needs search or deeper reasoning, Voice hands off to a text model: since September 23, 2026 paid users can pick GPT-6 Astra ($10/$50 per million tokens on the API), GPT-6 Sol ($2/$10) or GPT-6 Luna ($0.10/$0.50), which OpenAI released on September 3 and September 22 respectively. Voice inside ChatGPT Work follows the plan's Voice limits and draws on normal Work usage.

What are the ChatGPT Voice usage limits in 2026?

Live usage is measured over a rolling 24-hour window. Free gets limited access to GPT-Live-1 mini; Go gets up to 3 hours on GPT-Live-1 mini; Plus gets up to 3 hours on GPT-Live-1; Pro at $100 per month gets 15 hours; Pro at $200 per month gets unlimited GPT-Live-1. Business Standard gets 3 hours with extra usage at 1.25 credits per minute, Business Premium gets 15 hours, Enterprise, Edu and Clinicians on credit pricing pay 1.25 credits per minute, and Enterprise on usage-based pricing pays $0.05 per minute.

Can ChatGPT Voice approve actions by voice?

No. When a plugin action needs approval on web or mobile, Voice prompts an on-screen review with buttons to approve or decline, and OpenAI states that approving by voice is not supported. Existing app connections, permissions and workspace restrictions carry over unchanged. Two other limits as of September 2026: Live mode does not support video or screen sharing (the older Advanced voice mode still does on mobile), and Live cannot find or add files from your ChatGPT Library, though you can attach a file manually.

Quick Answer

What's New in ChatGPT Voice: Plugins, Work, GPT-6 Switching

Published: September 25, 2026

What OpenAI shipped on September 23, 2026

OpenAI’s ChatGPT release notes for September 23 record the largest change to Voice since GPT-Live-1 arrived in July. Voice stops being a conversation feature and becomes a hands-free front end to everything else in ChatGPT: plugins, connected apps, the Work agent, and the GPT-6 model family.

Change	What it means
Plugins in Live voice (web, iOS, Android)	Ask ChatGPT to use any plugin or connected app your account already has during a spoken conversation; the written response appears in the chat. Free and Go users get the plugins their plan supports.
Voice in ChatGPT Work (web, mobile)	Start Work tasks by speaking: draft documents, presentations and spreadsheets, draw on connected apps, or operate in a browser. If a task is still running when you hang up, it continues in text. Requires both Voice and Work access.
GPT-6 model selection	Paid users pick GPT-6 Astra, Sol or Luna as the reasoning backend Voice escalates to. Previously Voice was tied to a single handoff model.
Voice in the desktop app (macOS, Windows)	Start tasks, check progress, ask questions about your agents and coordinate multiple agents from one spoken conversation; this had been available since July 23 and is now documented alongside the mobile changes.

How ChatGPT Voice is structured

OpenAI’s help article lists three Voice options, and the distinction matters for what you can do:

Live runs on GPT-Live-1 (paid) or GPT-Live-1 mini (Free). It listens and speaks simultaneously, uses web search and memory, shows visual results in supported widgets and works with text and images in the same chat. This is where plugins and Work landed on September 23. It does not support video or screen sharing, and it cannot pull files from your Library.
Advanced is the previous real-time mode, kept because it still supports video and screen sharing on mobile.
Standard is turn-by-turn: it transcribes your speech, then responds.

Under the hood, when Live needs to search or reason it hands off to a text model. That handoff first became configurable on September 9, 2026 (GPT-5.6 or GPT-6 Astra); the September 23 update opens it to the full GPT-6 lineup: Astra for the hardest questions, Sol for everyday professional work at a fifth of Astra’s API price, Luna for fast high-volume answers.

Limits and billing

Live usage is metered over a rolling 24-hour window:

Plan	Live allowance
Free	Limited GPT-Live-1 mini
Go	Up to 3 hours, GPT-Live-1 mini
Plus	Up to 3 hours, GPT-Live-1
Pro ($100/mo)	Up to 15 hours
Pro ($200/mo)	Unlimited GPT-Live-1
Business Standard	3 hours, then 1.25 credits/min
Business Premium	15 hours
Enterprise / Edu / Clinicians (credits)	1.25 credits/min
Enterprise (usage-based)	$0.05/min

Voice inside Work follows the same Voice limits and consumes normal Work usage exactly as typed tasks do.

The approval rule

The most important design decision in the update is the one OpenAI put in a single sentence: approving by voice is not supported. When a plugin action needs consent, Voice raises an on-screen review with approve and decline controls, and existing app permissions and workspace restrictions apply unchanged. That closes the obvious attack, a spoofed or misheard “yes” authorizing an email or a purchase, and it is the same posture Google took with Call for Me, where a live transcript and a takeover button gate every call. Anyone building voice agents on the OpenAI Agents API should copy the pattern.

How it got here

July 8, 2026: GPT-Live-1 and GPT-Live-1 mini launch in ChatGPT Voice with simultaneous listen-and-speak, web search, memory and widgets.
July 9, 2026: ChatGPT Work launches as the long-running agent for documents, spreadsheets, decks, reports and Sites, on paid plans.
July 23, 2026: Voice arrives in Work and Codex inside the desktop app.
August 31, 2026: Live conversations appear on the iPhone Lock Screen and Dynamic Island via Live Activities.
September 9, 2026: Voice can hand off to GPT-5.6 or GPT-6 Astra for hard questions; daily limits simplified.
September 23, 2026: plugins in Live, Voice in Work on web and mobile, GPT-6 Astra/Sol/Luna selection.

What it does not do

It does not make phone calls, it does not see your screen in Live mode, and it does not approve anything you say out loud. For the voice-model comparison against Google, see Gemini 3.8 Live vs Extended Thinking vs GPT-Live-1.

Last verified: September 25, 2026.