Quick answer

On September 23, 2026 OpenAI added plugins to ChatGPT Voice: connect email, calendar, and Slack, and a voice conversation can read, summarise, draft, and schedule. The same update lets you choose which GPT-6 model (Astra, Sol, or Luna) powers voice. Alongside it, ChatGPT's mobile app gained Work agents that run multi-step tasks. Each connection asks for account permissions; read them before granting.

Voice assistants were always a step behind because they could not do anything. This update closes part of that gap: talk to ChatGPT on the walk to work and it can tell you what is in your inbox, move a meeting, and post to a channel.

What the plugins do

  • Email: summarise unread mail, search, draft and send replies by voice
  • Calendar: read your day, find free slots, create and move events
  • Slack: catch up on channels, draft messages, post to a channel with confirmation
  • Model choice: run voice on Astra for agentic tasks, or Sol and Luna for cheaper everyday use

Before you connect anything

Each plugin requests access to an account, and the practical question is scope: read-only or read-and-send, one calendar or all, which Slack workspaces. Work accounts may be governed by your employer's policies on third-party apps; check before connecting a corporate mailbox. OpenAI's own guidance recommends confirmation steps for send and post actions; leave them on. And the September 27 disclosure that a separate OpenAI research agent leaked user images is a reminder that connected data is only as safe as the weakest feature holding it.

How it compares

Meta Muse, launched September 8, goes further: it runs errands and fills forms inside a secure virtual machine with its own browser. ChatGPT's approach is narrower and, for now, safer: specific plugins with specific permissions rather than a general-purpose agent with your logins. Which is better depends on whether you want an assistant that reads your inbox or one that books your flights.

Bottom line

Voice that can act is genuinely useful, and the permission model is sensible. Connect a personal account first, keep confirmations on, and decide about the work mailbox after you have seen how it behaves.