DeskTheory is where founder-CEOs learn to run their companies on AI leverage.

What is GPT-Live?

GPT-Live is OpenAI's family of voice models that speak and listen at the same time, the way a person on a phone call does. It replaced ChatGPT's old voice mode on July 8, 2026, and it's what powers ChatGPT Voice on your phone and desktop today.

You'll hear this thing called GPT-Live, ChatGPT Voice, or just "the new voice mode." Same thing. GPT-Live is the model family; ChatGPT Voice is the feature it powers. Either way, the test is simple: interrupt it mid-sentence. It stops, listens, and adjusts, because it was listening the whole time it was talking.

What it is (in plain English)

Every voice AI you've used before worked like a walkie-talkie: you talk, it waits, it talks, you wait. The models could only process one side of the conversation at a time, which is why cutting one off always felt like a glitch.

GPT-Live is full-duplex, meaning it processes speech in and speech out simultaneously. That one architectural change is what makes it feel like a phone call instead of a walkie-talkie: natural interruptions, back-and-forth at conversation speed, even live translation between two people talking over each other.

OpenAI shipped it on July 8 in two sizes: GPT-Live-1, the default on paid ChatGPT plans, and a mini version covering the free tier, across iOS, Android, and web. On July 23 it landed in the ChatGPT desktop app with a programmable hotkey, and this is the version that changes the job: talk through what you need and ChatGPT starts moving several tasks forward while you keep working, including code through Codex. In August, voice conversations picked up file uploads and Projects support, so you can hand it a document mid-call and ask about it.

One more piece worth knowing exists: since July 31, audio generated by GPT-Live carries a SynthID watermark, and OpenAI runs a public tool that can verify whether a clip came from its models. In a year of convincing fake audio, provenance checking is now a thing you can actually do.

Why you should care as a CEO

Pretty much everyone talks faster than they type, and your working day is already spent talking. Voice is the fastest way to hand work to an AI, and full-duplex is what finally makes it feel like handing work to a person. The old voice mode was a demo you showed your kids; this one is an input device.

The desktop hotkey is the operator move. You're in the spreadsheet, you hit the key, you say "pull together what we know about the Meridian renewal and draft my reply to their CFO," and you keep working while it does. Dictate the brief while walking to the next meeting. Tell it to prep the 2pm while you finish the 1pm. The pattern we teach in voice mode for CEOs just got a much better engine, and your phone got it first.

The catch is the same as every voice interface: it's only as useful as what it's connected to. Voice plus your calendar, email, and files is a chief of staff; voice alone is a very pleasant search box.

Where you'll see it

• In ChatGPT on your phone and in the desktop app, anywhere you see the voice icon or set the hotkey.
• In our news desk coverage: ChatGPT finally talks like a person, and it works while you talk.
• In voice mode for CEOs, the workflow this model upgrades.
• In the deepfake conversation, where SynthID watermarking on generated audio is becoming the verification layer.

What you should do next

Set the voice hotkey in the ChatGPT desktop app and delegate the next thing you were about to type; interrupt it rudely at least once, since that's the feature.