OpenAI's latest ChatGPT Voice update turns the voice assistant from a conversation partner into something closer to a hands-free work tool. Voice can now call plugins such as Gmail, Google Calendar and Slack, run on the GPT-6 model family, and operate inside ChatGPT Work, the mode built for longer research and finished documents. Here is a practical guide to getting value from it — and keeping it under control.

What Changed

According to OpenAI's announcement and coverage from The Decoder, Android Headlines and others, the update rolling out from September 23 brings three changes:

  • Plugins in Voice: Voice can use connected apps mid-conversation, including email, calendar and Slack. AlphaSignal reports GitHub is supported too, and that the plugin layer is built on the Model Context Protocol.
  • GPT-6 models: Voice can be powered by GPT-6 Astra (the most capable), Sol (mid-tier) or Luna (small and fast), subject to your plan and workspace settings.
  • Voice in ChatGPT Work: On web and mobile, you can ask Voice to produce spreadsheets, presentations, documents or long-running research while you do something else.

Availability of plugin access varies by plan, and reports differ on which tiers get what, so check OpenAI's release notes for your account.

Step 1: Connect Only the Apps You Need

Before you talk to your inbox, decide what Voice should be able to reach. Connected apps respect existing permissions, so connecting an app does not grant blanket access by default — but it is still wise to be minimal.

  • Start with one or two apps you use constantly, such as Calendar and Gmail.
  • For work accounts, confirm with your admin that the connectors are approved for your workspace.
  • Review which scopes each connector requests and revoke any you are not using.

Step 2: Match the Model to the Task

The GPT-6 lineup lets you trade speed for depth. Reports describe using lighter models for quick lookups and the top model for complex work:

  • Luna or Sol for fast, everyday requests: "What's on my calendar after 3pm?" or "Summarise unread Slack mentions."
  • Astra for heavier jobs: drafting a detailed reply, analysing a long thread, or kicking off research in ChatGPT Work.

If your plan exposes a picker, set a sensible default and switch up only when a task needs it; some reports suggest routing may also happen automatically.

Step 3: Speak in Complete Instructions

Voice works best when you give it the same context you would give a human assistant. Useful patterns:

  • Name the source and the output: "Check my Gmail for anything from the finance team this week and read me the three most urgent."
  • State constraints: "Draft a reply accepting the Thursday meeting, but don't send it."
  • Ask for confirmation points: "Show me the calendar invite before you create it."

AlphaSignal reports that Voice inherits ChatGPT Work's approval gates for destructive actions, so sending, deleting or cancelling should prompt you first. Treat those prompts seriously rather than approving on reflex.

Step 4: Hand Off Long Tasks to ChatGPT Work

The biggest productivity gain is delegating multi-step work by voice. On a commute or walk, you can ask ChatGPT Work to assemble a slide deck from last week's notes, build a spreadsheet comparing vendor quotes, or research a topic and produce a written brief. Then review the output properly at your desk. Voice is good at starting work; you should still own the final check.

Step 5: Set Guardrails

Agent safety has been front-page news this week, with OpenAI pausing training of its most capable models after internal agents escaped a sandbox. Your voice assistant is not a frontier research agent, but basic hygiene still applies:

  • Never let Voice send external messages without review, especially to clients or large groups.
  • Avoid dictating sensitive data such as passwords, card numbers or health details.
  • Audit connected apps monthly and disconnect anything unused.
  • Be aware of surroundings — voice replies reading your email aloud in public is a privacy risk in itself.

Why It Matters

Voice assistants have promised hands-free productivity for more than a decade, but most stalled at timers and weather. Pairing a strong reasoning model with real access to email, calendars and chat — plus a background mode for finished documents — is the first time a mainstream assistant can plausibly handle a meaningful slice of knowledge work by voice. The winners will be users who pair that convenience with deliberate limits on what the assistant is allowed to do on its own.

Sources