Hello everyone,

For two years, talking to ChatGPT has felt like using a walkie-talkie. You speak, you wait, it responds, you wait again. On July 8, OpenAI finally killed that rhythm.

The old Advanced Voice Mode chained three separate models together — one to transcribe your speech, one to think of a reply, one to turn that reply back into audio. Every handoff added lag, and the system could only listen or talk, never both, which is why it kept talking over you the moment you paused to think.

GPT-Live, OpenAI's new voice model, throws that pipeline out. It's full-duplex — it processes what you're saying while it's still speaking, the way two humans manage a real conversation. It knows the difference between finishing a thought and just thinking, and uses small verbal cues, a quick "mhmm" or "got it," to signal it's still listening.

Paid subscribers on Go, Plus, and Pro get the full GPT-Live-1 model. Free users get the smaller GPT-Live-1 mini — still full-duplex, just less capable. Both are rolling out now on chatgpt.com, iOS, and Android, with an API version expected within days. Business, Enterprise, and Edu accounts aren't included yet, and video or screen sharing still requires the old Advanced Voice Mode.

Voice has always been the most natural way humans communicate, and the clunkiest way to talk to AI. This update closes a chunk of that gap in one release.

Sign up today

Here's the detail that matters more than the natural cadence: GPT-Live doesn't try to be smart on its own. Ask it something that needs a real answer — a web search, a calculation, actual reasoning — and it quietly hands the question to GPT-5.5 in the background, then folds the answer back into the conversation. OpenAI says that's why GPT-Live-1 scores 75.2% on the BrowseComp web-search benchmark, versus 0.7% for the old voice system. That's not an incremental bump — that's the difference between an assistant that can look things up and one that basically couldn't.

That delegation trick is the real strategic move, not the natural-sounding voice. OpenAI is decoupling the personality layer from the intelligence layer — the voice model stays light and fast, while reasoning gets handed off to whatever the current best model is. Every time GPT-5.5 gets replaced, GPT-Live gets smarter without retraining from scratch — the same "small model routes to big model" pattern showing up across the industry.

Voice is also the interface where AI safety concerns feel most immediate. OpenAI says it expanded audio-specific safety evaluations for self-harm and emotional reliance, added real-time safeguards that can redirect a conversation mid-sentence, and built in parental controls that notify a linked parent if a teen's conversation signals risk. Whether that holds up outside a controlled demo is the real test — text-based safety systems have had plenty of failure modes of their own.

If you write, plan, or think out loud for a living, this is worth trying today. A voice assistant that can be interrupted, that delegates hard questions instead of bluffing, and that can pull up a chart or a map mid-conversation starts to look less like a novelty and more like something you'd build a workflow around.

Voice just stopped being a convenience feature and became a genuine second interface to the same underlying intelligence. Expect Google and Amazon's assistants, already moving this direction, to face real pressure to match the delegation trick, not just the vibe.

Have you tried GPT-Live yet? Hit reply and tell us if it actually feels different. And send this to the one friend who still refuses to use voice mode.

See you next time.