TL;DR
- Since September 23, you can give ChatGPT a multi-step job by voice inside Work, its space for longer tasks, and if you hang up before it's done, it keeps going in text.
- The same day, Google opened voice replication: from a clip of your voice up to 30 seconds long, plus a consent line you record yourself, it builds a synthetic voice that sounds like you, marked as AI-made.
- And a PDF detail: Claude reads both the text and the images on every page, which adds up; ChatGPT, outside Enterprise accounts, pulls only the text, so a scan can reach it nearly empty.
- Further down, two things to copy: a line that lets you dictate a task without ChatGPT jumping in mid-thought, and a script for recording your voice in one take.
Some of the best thinking of the day happens when you can't type: walking between appointments, waiting before a meeting. Whatever came to mind there usually got lost.
What changed on the 23rd is that those minutes can now produce a draft.
ChatGPT: hand it the job out loud, and it keeps working after you hang up
What changed. On September 23, OpenAI's release notes said ChatGPT's voice now uses the apps connected to your account and works inside Work, ChatGPT's space for multi-step jobs: building a document, a deck or a spreadsheet, using your connected apps, or working in the browser. The help page puts it plainly: if a task is still running when you end the call, "it can continue in text."
One limit from the same page: anything that needs your approval gets approved or declined with the on-screen controls, because "spoken approval is not supported."
Who has it. Work is on the paid plans that include it, on web and mobile. If you don't see Work in the menu, your plan doesn't include it, or on a company account, your admin hasn't turned it on.
Where you click. OpenAI's help page gives the path: open ChatGPT, choose Work, then select the Voice control to start. On mobile, pick Work from the dropdown at the top of the screen.
If you're driving, set it up before you pull out. OpenAI asks you not to handle the phone while the car is moving, and doesn't document Work in CarPlay. Save the approvals for when you're parked.
⚠ These steps come from OpenAI's help pages, not from our own run. Nobody on our team has dictated a Work task yet. The order below is ours.
How to dictate a task
- Say the result before the topic. "I need an email ready to send" works better than "let's talk about yesterday's client."
- Say what it should work from. If it needs your email or calendar, name it: that app has to be connected, and it can't see what isn't.
- Say what it can't do without you. Send, delete, book. Say it up front, and it won't build the job around a send you didn't want.
- Hang up and get on with your day. If the task isn't finished, you'll find it in writing in the same conversation.
The line to dictate
The brackets are yours to fill. The rest stays the same every time:
"Don't answer until I say 'done.' I need [an email, a summary, a one-page proposal] for [who]. Use [what it should check: this person's last email, my calendar for tomorrow]. What has to come across is [the decision or the fact that matters]. Nothing gets sent or booked: I want it in writing. Done."
When you think out loud you pause, and ChatGPT's voice can take any pause as its turn and start on half the request. OpenAI suggests exactly this: ask it up front to wait until you tell it to answer.
Google: a synthetic voice that sounds like you, if you authorize it on the recording
What changed. On September 23, Google introduced Gemini 3.8 Flash TTS and with it voice replication. The docs ask for a 10 to 30 second reference clip and for the same adult to read a fixed consent statement, available in 30 locales including U.S. Spanish. The system compares the two recordings before it builds the voice. Every clip carries Google's SynthID watermark - a signal you can't hear - and C2PA content credentials, a record that travels with the file and says how it was made.
Where you click. In Google AI Studio, on the speech page: aistudio.google.com/generate-speech. Per Google's docs, that's where you record or upload both clips in the browser. We couldn't see the menu inside, so we're not giving you button names we didn't see.
Where it isn't. Voice replication in AI Studio is not available in Illinois, Texas, the EEA, the U.K., Switzerland or India. Florida is fine.
What it's good for in a service business
Narrating what you've already written: the audio version of your article, the voiceover on a short video, the explainer for a service you re-record every time a line changes. Change a line, and the voice reads the new version.
The script to record your voice in one take
Twenty to twenty-five seconds, read calmly, in the tone you use with a client, in a room without echo:
"Good morning. Here's how we work, in a few words. First we listen to what you need and what you've already tried. Then we tell you plainly what we can do, how long it takes and what it costs. And if we're not the best fit for you, we'll tell you that too. We'd rather lose a sale than win a client we can't serve well."
Four more worth a look
We checked each one on its own site on September 28. No paid placements here.
- SaneBox - moves the unimportant mail out of your inbox, in whatever email you already use. Seven-day free trial, no card.
- Schole - builds short AI lessons around each person's role on your team. Free to start.
- Naise - runs social, influencer and PR campaigns from a single written brief. Three-day trial, card required.
- Gemini Notebook - what used to be called NotebookLM: upload your own documents and ask questions about them. Has a free plan.
Three things the announcement leaves out
The same PDF doesn't reach two assistants the same way. Per Anthropic's help center, in a chat Claude reads both the text and the images in PDFs up to 100 pages. Its technical docs explain how: each page becomes an image, and each page costs 1,500 to 3,000 tokens of text plus whatever the image costs. ChatGPT does the opposite: per its help pages, on every plan except Enterprise it pulls only the text from a PDF and drops the images, and it warns that with a scan it may not extract the data well.
It's the same call we made when we set up our visual system in Claude Design: there the PDF is the point, because it has to look like your brand.
If someone helps you with your computer, ask them for MarkItDown once. It's a free, open-source Microsoft tool that turns PDF, Word, Excel and PowerPoint files into clean text. It runs from the command line, so it isn't for you: it's for that person to hand you your five most-used documents as text.
Voice doesn't approve. That's a safeguard: in ChatGPT, nothing that needs approval goes out on something said out loud. And the same help page admits voice sometimes answers people talking nearby, not you.
What they do badly
Voice inside Work draws from two places at once: your voice limit, which on Go and Plus is three hours per rolling 24, and your Work usage, same as if you typed. Add it to the tally of what AI already costs you each month: no new bill shows up here, what you already pay just runs out faster. And it only sees what's connected: if your work email isn't, the task can't read it, and the help page doesn't say it will warn you.
Voice replication depends on the recording you give it: a bad clip makes a bad voice. And the docs don't say a voice recorded in Spanish will sound right in English, so we won't promise it. And AI Studio is built for people building with Gemini: it runs in a browser, but it isn't a simple app.
Who should skip it
If you've never asked ChatGPT for anything longer than an answer, start with voice in regular chat. If your business doesn't publish audio or video, leave voice replication for later.
Where it fits in your business
In a practice or a firm, what's left at the end of the day is rarely hard: it's writing up what was already decided, like the follow-up email or the client summary. Dictate it between appointments, and it's a draft by the time you're back at the office.
Before you upload a PDF with client details, remember that turning that setting off doesn't erase what you already typed.
The one thing to do today
Keep reading
- Meta put a price on the AI you were already using - last Tuesday's tools
- Turning that setting off doesn't erase what you already typed - what client data never goes into an AI tool
- Claude Design has a setting most people never turn on - how we built ours
- SEO vs. GEO vs. AEO - the whole map
- ChatGPT Voice, OpenAI help center - the source
Where to start
The voice features depend on your plan and where you live. The text version of that document depends on nothing: no paid plan, no new app.
Sources
- ChatGPT release notes, OpenAI, September 23, 2026 - https://help.openai.com/en/articles/6825453-chatgpt-release-notes
- ChatGPT Voice, OpenAI help center, checked September 28, 2026 - https://help.openai.com/en/articles/20001274-chatgpt-voice
- ChatGPT Work and Codex, OpenAI help center, checked September 28, 2026 - https://help.openai.com/en/articles/20001275-chatgpt-work-and-codex
- File uploads FAQ, OpenAI help center, checked September 28, 2026 - https://help.openai.com/en/articles/8555545-file-uploads-faq
- Gemini 3.8 Flash TTS, Google, September 23, 2026 - https://blog.google/innovation-and-ai/models-and-research/gemini-models/gemini-3-8-text-to-speech/
- Voice replication, Gemini API docs, Google, checked September 28, 2026 - https://ai.google.dev/gemini-api/docs/voice-replication
- Upload files to Claude, Anthropic help center, checked September 28, 2026 - https://support.claude.com/en/articles/8241126-upload-files-to-claude
- PDF support, Claude docs, Anthropic, checked September 28, 2026 - https://platform.claude.com/docs/en/build-with-claude/pdf-support
- MarkItDown, Microsoft, checked September 28, 2026 - https://github.com/microsoft/markitdown
