ChatGPT Voice Adds Plugins to Live and Expands Work Access: Where It Works and How Usage Is Counted

OpenAI’s Voice update makes connected tools available during Live conversations and expands access to Voice in ChatGPT Work. The practical details depend on where you start the conversation: Chat, Work, and the desktop app have different model access and usage rules. GPT-6 Sol and Luna are documented for Work and Codex; their role in Voice conversations in Chat needs a more specific source.

The update shipped on September 23, 2026, globally in the latest app version. Below: what changed on each surface, what did not, how usage is counted in each, how approvals and privacy work, and who benefits. We have not tested the new features ourselves.

The update at a glance

• Plugins in Live: on web, iOS, and Android; OpenAI names email, calendar, and Slack; other plugins follow your existing connections and permissions
• Voice in ChatGPT Work: new on web and mobile; create documents, presentations, spreadsheets, and sites, use connected apps, or work in a browser; requires Voice and Work access
• Desktop app: Voice in Chat, Work, and Codex was already available before this update
• Models: GPT-Live-1 or GPT-Live-1 mini runs the conversation; a second model handles reasoning and tools. Documented for Chat: GPT-5.6 and GPT-6 Astra. Documented for Work and Codex: GPT-6 Sol and Luna, plus Astra
• Approvals: on screen only
• Unchanged: Live has no video or screen sharing; Advanced and Standard remain as options

What changed, by surface

Voice in Chat: plugins inside a Live conversation

During a Live conversation you can ask ChatGPT to use a plugin available to your account. OpenAI’s announcement names email, calendar, and Slack; the help article extends this to the plugins and connected apps your account already has, with the same permissions and workspace restrictions as typed chat. You may need to connect an app and sign in first. Written responses appear in the chat alongside the spoken ones. Whether custom connectors are covered is not stated.

The boundary is the approval step. Actions a plugin flags as needing approval are shown on screen, and you approve or decline with the on-screen controls; spoken approval is not supported. Actions within your existing permissions can run without a prompt, so the behavior depends on how each plugin’s permission level is set. Free and Go users get plugins in Voice within what their plan supports. Our ChatGPT Voice mode guide covers the conversational basics.

Voice in Work: now on web and mobile

ChatGPT Work is OpenAI’s agent surface, separate from Chat. Voice in Work lets you ask for documents, presentations, spreadsheets, and sites, use connected apps, or drive a browser by speaking, and an unfinished task continues in text after the call ends. What is new on September 23 is the web and mobile access; the desktop app already supported Voice across Chat, Work, and Codex, including starting tasks, checking progress, and switching between Codex tasks by voice. Both Voice and Work access are required.

Which models sit behind Voice

ChatGPT Voice runs on two layers. GPT-Live-1, or GPT-Live-1 mini on Free and Go, handles the real-time conversation. When a question needs search, reasoning, or a tool, Live hands it to a second model and keeps talking while that model works. Our GPT-Live-1 explainer describes the design.

What is documented for that second model differs by surface. For Voice in Chat, OpenAI’s September 9, 2026 release note lists GPT-5.6 and GPT-6 Astra, chosen with the same model and effort controls as text chat; the older Instant, Medium, and High voice levels were retired at the same time. For Work and Codex, the models page lists GPT-6 Astra, Sol, and Luna, and states that Sol and Luna are not available in Chat. OpenAI’s announcement says Voice can be powered by all three GPT-6 models without naming a surface; the documented reading is Sol and Luna behind Voice in Work and Codex, and Astra behind Voice anywhere your plan allows. Our GPT-6 Sol and Luna piece covers those models.

What did not change

Live still does not support video or screen sharing; those remain in Advanced on iOS and Android. Standard remains as a turn-by-turn option. Preset personalities do not apply to Live. Live is built for one speaker. All three modes stay selectable under Settings and Voice, subject to plan, region, and workspace settings.

How usage is counted

Voice in Chat has its own allowance, measured over a rolling 24-hour period.

Source: OpenAI Help Center, ChatGPT Voice, September 23, 2026.

Voice in Work on web and mobile follows the same Voice allowance and billing, and the Work task itself counts against your Work usage exactly as a typed task would. Voice in the desktop app is priced separately: on personal plans it draws $0.05 per minute from the Codex usage budget; on Business, Edu, and Enterprise workspaces with credit-based billing it costs 1.25 credits per minute; in both cases the model doing the task is billed at its own token rates on top. A call also ends at the maximum session length or when a long conversation hits its context limit.

Who should use it

Voice with plugins fits situations where you can glance at a screen: walking, standing in a queue, cooking. It does not fit driving, because any action that needs approval waits for a tap, and OpenAI’s CarPlay guidance is to set up before driving and not interact while the vehicle is moving. Good uses include asking what is in your inbox or calendar and approving replies on screen, dictating a document or slide outline into Work to edit later, and turning spoken notes into a summary posted to Slack. Tasks that need precise placement, such as which spreadsheet cell to change, remain easier to point at than to describe, and the transcript in your chat is not a verbatim record.

Before connecting a work account, review each plugin’s permission level under Settings and Plugins: Live acts within what you have already granted, and OpenAI’s global plugins permission setting defaults to allowing low-risk actions without asking.

Privacy controls, by data type and account

• Audio and video clips. Clips from Live and Advanced are stored with the transcript and retained for 30 days; deleting the chat deletes them within 30 days, archiving does not. Standard deletes audio right after transcription. OpenAI does not train on clips unless a personal account turns on Improve the model for everyone under Settings and Data Controls and then enables Include your audio recordings or Include your video recordings.
• Transcripts. If Improve the model for everyone is on, transcripts and files from Voice conversations may be used for training, depending on plan and settings.
• Business, Enterprise, and Edu accounts. Users cannot share audio or video clips for training at all; workspace settings govern the rest.

The source for all of this is the data controls section of OpenAI’s ChatGPT Voice help article, listed above.

Documented, not documented

FAQ