OpenAI Launches GPT-Live Voice Model Worldwide, Flags Its Own Emotional Risk

OpenAI has switched on GPT-Live, a voice model that lets ChatGPT listen and talk at the same time, for users around the world. The company calls it its smartest voice model yet, built to replace the two-year-old Advanced Voice Mode with something that interrupts less and knows more.

Five days into the rollout, the feature drawing the most praise, natural backchannel sounds like “mhmm” that make the model feel like it is really listening, is also the one OpenAI’s own safety card flags as a step back. GPT-Live-1 scored 0.82 on an internal emotional reliance measure, down from 0.88 for Advanced Voice Mode. OpenAI says the dip is not statistically significant.

Two Models Quietly Retire a Two-Year-Old System

ChatGPT’s original voice tool, launched in 2023, chained three separate models together. Whisper turned speech into text, GPT-4 wrote a reply, and a text-to-speech engine spoke it back. It worked, but the relay added lag and flattened tone.

Advanced Voice Mode replaced that pipeline in 2024 with a single audio-native model, cutting latency and preserving more of how something is said. It still made people wait for a clean pause before answering, and a cough or a stray “um” could trigger an interruption at the wrong moment.

GPT-Live removes the wait. The model processes incoming audio and generates speech at the same time, and can decide many times a second whether to talk, listen, pause or interrupt. Two versions shipped globally on July 8: GPT-Live-1 for paid tiers and a lighter GPT-Live-1 mini for the free tier, both replacing Advanced Voice Mode by default across iOS, Android and the web.

More than 150 million people already use ChatGPT’s voice and dictation features every week, according to OpenAI, which is the base GPT-Live inherits on day one.

What Changes the Moment You Start Talking?

GPT-Live lets a user talk over ChatGPT mid-reply, pause without losing the floor, or ask it to slow down, and the model answers back with quick acknowledgements like “mhmm” and “got it.” It also picks a reasoning level, filters background noise, and can show a visual card instead of just speaking an answer.

• Natural conversations – interrupt mid-reply, pause to think, or ask ChatGPT to slow down, with quick verbal acknowledgements to show it is following along.
• Smarter responses – three reasoning levels, Instant, Medium and High, trade speed for depth on demand.
• Better listening – the model waits through pauses, stays silent on request, and filters out background noise like traffic.
• Visual answers – weather, stock and sports questions can now surface as on-screen cards instead of speech alone.

OpenAI also re-recorded all nine ChatGPT voices for the new model. Because ChatGPT already works inside Apple’s CarPlay, the upgrade reaches car dashboards too, letting drivers redirect a conversation hands free, iThinkDifferent reported.

Free Users Get the Smaller Brain

The rollout splits cleanly by plan.

• Free – GPT-Live-1 mini, Instant reasoning only.
• Go, Plus, Pro – GPT-Live-1, with a choice of Instant, Medium or High reasoning.
• Business, Enterprise, Edu – not available at launch; these workspaces stay on Advanced Voice Mode.

Real-time voice sessions also cannot pull from ChatGPT’s saved memory yet, BigGo Finance reported, though OpenAI says that gap is temporary. Video calls and screen sharing are missing too, and both still require the legacy voice mode for now.

Language support is uneven. OpenAI optimized the model for its most-used languages and warned that others may carry a non-native accent or fluency gaps. That limit showed up during a press demo, when a live Hindi translation came out sounding stiff and heavily accented, according to TechCrunch.

The Numbers Behind OpenAI’s Smartest Voice Model Claim

OpenAI built new evaluations just to measure how pleasant a conversation feels, since existing voice benchmarks were not designed for a model that talks and listens simultaneously. Testers judged matched five-to-ten-minute conversations on turn-taking, interruptions, flow and how natural each interaction felt.

The clearest gains show up in reasoning and search. Figures compiled by outlets including The Decoder and MLQ AI put GPT-Live-1 at 84.2% on GPQA, a graduate-level science reasoning test, at its highest reasoning setting, nearly double Advanced Voice Mode’s 45.3%. On BrowseComp, a benchmark for tracking down hard-to-find information online, GPT-Live-1 reportedly jumped to 75.2% from 0.7%.

OpenAI told EdTech Innovation Hub that testers preferred GPT-Live-1 over the old system 75.7% of the time, and GPT-Live-1 mini 69.2% of the time.

The company is betting the gains extend beyond a nicer chat. Atty Eleti, OpenAI’s product lead for ChatGPT Voice, described taking long walks while working through problems out loud with the model.

Over time, we think this will also unlock the ability to use voice as a kind of primary interface to computing, and to manage increasingly complex long-running agentic work.

Eleti said that during a press briefing, arguing voice could become the main way people direct long, complex AI work rather than a shortcut for quick questions.

The Selling Point Doubles as a Warning Sign

GPT-Live’s backchannel, the soft “mhmm” and “yeah” that signal it is listening, is the detail OpenAI leads with in its own marketing. It is also, per outside research, the behavior closest to what has been linked to emotional dependency.

GPT-Live-1’s emotional reliance score slipped from 0.88 to 0.82, a dip OpenAI’s safety team calls statistically insignificant. GPT-Live-1 mini showed a smaller slip too, from 0.97 to 0.95 on a separate sexual content measure.

OpenAI built dedicated safety training across self-harm, psychosis and mania, emotional reliance, violence and sexual content for the launch. Real-time safeguards can steer a reply mid-sentence, surface crisis resources, or end a session entirely in high-risk cases.

That caution has research behind it. A 2025 joint study between OpenAI and MIT Media Lab, covering nearly 1,000 participants and more than 3 million ChatGPT conversations, found that heavy Advanced Voice Mode users who discussed personal topics showed more emotional dependence and less real-world socializing over time, Tech Times reported. A separate April 2026 STAT News analysis, cited by Tech Times, went further, tying extended voice use to a mental health crisis that included a Florida man’s death by suicide after months of interaction.

OpenAI acknowledges the tension. The company says it is rolling out longer-term monitoring focused specifically on emotional reliance now that GPT-Live is designed to hold attention for longer than earlier voice sessions typically did.

Early Reaction Splits Between Awe and Annoyance

A launch post on X, tracked by the account Digg, drew more than 500,000 views with mostly positive reaction, though the post itself questioned whether voice AI can finally stop feeling like a demo.

Not everyone is charmed. BigGo Finance reported social media complaints calling the model’s constant “mhmm” and “yeah” interjections annoying, with some users saying the acknowledgements feel intrusive rather than reassuring.

On Hacker News, one developer welcomed the GPT-5.5 delegation but wrote that the model was “interrupting me and laughing at my (not really intended as) jokes while I was still talking,” before OpenAI toned the behavior down.

Writer Aadit Sheth described a different edge case: a restaurant’s reservation confirmation call that turned out to be AI. Once he began testing it with questions, in his words, it just hung up on me.

A Crowded Field Is Chasing the Same Bet

OpenAI is not the only lab chasing full-duplex voice. Google’s Gemini Live already handles simultaneous listening and speaking and adds camera and screen sharing, features GPT-Live does not yet have, according to BigGo Finance. NVIDIA shipped its own full-duplex model, PersonaPlex, earlier this year.

Anthropic rolled out a voice mode for Claude last year, and the rivalry between the two labs has only sharpened since, with Anthropic’s paper valuation recently jumping past OpenAI’s own after a fresh funding round. Apple and Amazon have both pushed their assistants toward more conversational, context-aware replies too.

The timing carries a corporate subplot. OpenAI confidentially filed draft paperwork for a stock listing on June 8, according to Prism News, and the voice relaunch lands just weeks after a court ruling that cleared its runway toward an $852 billion IPO by rejecting Elon Musk’s attempt to block the company’s for-profit restructuring.

Developers, meanwhile, are still waiting. OpenAI has not said when GPT-Live reaches the API, only that it is coming. ChatGPT Business, Enterprise and Edu workspaces do not have it at launch either, and video and screen sharing still require the older Advanced Voice Mode.

Frequently Asked Questions

Is GPT-Live Available Through the API Yet?

Not yet. OpenAI says access is planned but has not set a date, and developers can sign up through a notification form to be alerted when it opens.

Does GPT-Live Support Video Calls or Screen Sharing?

No, not at launch. Those features still require the legacy Advanced Voice Mode, which OpenAI says will keep running for accounts that need them until GPT-Live adds the capability.

What Is the Difference Between GPT-Live-1 and GPT-Live-1 Mini?

GPT-Live-1 is the larger model and the default for Go, Plus and Pro subscribers, with a choice of Instant, Medium or High reasoning. GPT-Live-1 mini is the default for Free accounts and runs Instant reasoning only.

Can Parents Turn off Voice Mode for Teen Accounts?

Yes. Parents can disable ChatGPT Voice for a teen entirely through Parental Controls, and OpenAI says a linked parent account gets a notification if a teen’s conversation shows signs of self-harm or suicidal intent.

Why Did OpenAI Retire Advanced Voice Mode?

Advanced Voice Mode still had to wait for a clean pause before replying, so background noise or a short pause could trigger an unwanted interruption. GPT-Live’s full-duplex design processes speech continuously, removing that wait entirely.