If you've tried ChatGPT's voice mode before and walked away feeling like you were talking into a walkie-talkie, the feeling was technically accurate.

The old system worked like a two-way radio: you talk, then it talks, then you wait for it to finish before you say anything else. Try to jump in mid-sentence and it either ignores you or stumbles into confusion. Fine for quick questions. Annoying for anything that takes more than thirty seconds to work through.

OpenAI changed that on July 8 with GPT-Live. The upgrade is about one thing: the model can now listen and speak at the same time.

What "full-duplex" means (and why it matters)

There's a term for what ChatGPT's voice mode can do now: full-duplex. It's how phone calls work. Both people can talk at once. You can interrupt. The other person hears you even if they were mid-sentence. The old voice mode was half-duplex — one direction at a time, like CB radio. "Over." "Over."

What changes in practice, for real people:

• You can say "wait, actually" in the middle of an answer and it will hear you. It won't barrel on for another thirty seconds ignoring the redirect.
• It can say "mhmm" or "yeah" while you're still talking, because it's already processing while you speak.
• The back-and-forth feels more like a conversation and less like submitting a form.

None of that is magic. It's what "can hear you while it's talking" enables. And it's the thing walkie-talkie voice mode made impossible.

Who gets what

Here's the practical part before you update your app.

Both tiers get the full-duplex architecture. That's the part that changes the feel. The difference between mini and the full GPT-Live-1 is what happens when you ask something hard: the full model delegates to a more powerful AI in the background and brings the answer back into the voice conversation when it's ready. On the free tier, GPT-5.5 handles everything directly.

In practice: both will feel noticeably better than old voice mode. Paid tiers will handle complex questions more gracefully.

Source spread

• OpenAI — Introducing GPT-Live — hype. The company's announcement. Clear on which tiers get which model, how delegation works for complex questions, and the rollout timeline for iOS/Android/web.
• TechCrunch — OpenAI releases new voice models — builder. Covers the full-duplex architecture and the GPT-5.5 backend detail independently.
• SiliconAngle — GPT-Live launch — builder. Notes context: GPT-Live ships ahead of the broader GPT-5.6 rollout; API access for developers is coming but not available yet.

What's real

• Full-duplex is a real architecture change, not a fine-tuning upgrade. If you've used voice mode before, you'll notice the difference in the first thirty seconds.
• Free users get it. GPT-Live-1 mini is the smaller model, but it's still full-duplex. The walkie-talkie problem is gone on all tiers.
• The rollout is global, across iOS, Android, and ChatGPT.com. If your app hasn't updated yet, it will soon.

What deserves a side-eye

• "Delegates complex questions to the frontier model" means latency comes back for hard questions. The snappy back-and-forth feel is for conversational exchanges. Ask it to write a detailed plan for something and you'll wait a bit.
• There's no published session length limit in the announcement. Worth testing before you rely on it for anything long.
• Developers who want to build with GPT-Live have to wait. API access is "coming soon," which is not a date.

What to do about it

• Update your ChatGPT app. GPT-Live is rolling out now to iOS and Android. Check the App Store or Google Play if the voice mode still feels like the old version.
• Try a real back-and-forth, not just a question. The full-duplex architecture shows most clearly when you're working through something interactively — a decision, a plan, a response to a difficult email. Try that use case, not a trivia question.
• If you're on Free, the mini model is still worth trying. Both tiers got the architecture upgrade. The paid tier handles harder questions better; the free tier handles normal conversation better than it used to.
• For long reasoning tasks, stick to text. If you need it to think through something complex and structured, the text interface is still better. Voice mode is for conversation, not for generating a ten-point plan.

Further reading

• OpenAI — Introducing GPT-Live — primary source; architecture, tiers, availability
• TechCrunch — OpenAI releases new voice models for more natural live conversations — independent coverage with technical context
• SiliconAngle — OpenAI launches GPT-Live voice model series — broader GPT-5.6 rollout context