OpenAI has replaced the walkie-talkie. On 8 July the company began rolling out GPT-Live-1 and GPT-Live-1 mini, a new family of voice models that can listen and speak simultaneously — full-duplex, in the industry’s term — ending the turn-based system that required users to stop talking before ChatGPT would respond, and that routinely mistook a pause for the end of a question.

The upgrade matters to marketers because voice is quietly becoming a serious interface for the interactions brands care about: support, product guidance, research and, increasingly, commerce. More than 150 million people already talk to ChatGPT through its voice and dictation features, by OpenAI’s count, and the quality of those conversations — whether the assistant interrupts, mishears or stalls — is starting to shape how customers judge any AI-fronted service. A voice mode that finally behaves like a phone call resets that baseline.

What GPT-Live-1 actually changes

GPT-Live-1 processes incoming audio while it is speaking, making interaction decisions many times per second. In practice that means users can interrupt mid-sentence, ask it to slow down, or leave it silent while they think; the model responds with brief acknowledgements — a “mhmm” or “got it” — to show it is following, and OpenAI says it is better at ignoring background noise. Kundan Kumar, OpenAI’s research lead for speech models, described the result as working “more like a phone call”.

The architecture also closes the intelligence gap that dogged the old voice mode. When a question needs reasoning, live web search or agentic work, GPT-Live-1 hands it to OpenAI’s frontier text models — currently GPT-5.5 — in the background while the conversation continues, weaving the answer back in when it lands. Some responses now arrive visually rather than spoken: weather forecasts, sports scores and stock charts appear as on-screen cards. Real-time translation ships too, streaming a spoken translation while the speaker is still talking, though the launch demo’s Hindi carried a heavy American accent — a reminder the localisation work is unfinished. The launch also brings nine remastered voices, and paid users can choose between intelligence levels, trading response speed against depth.

Who gets it, and what is missing

GPT-Live-1 mini is now the default voice model for free ChatGPT users, with the larger GPT-Live-1 reserved for paid Go, Plus and Pro plans; the rollout covers iOS, Android and the web over several days. Two gaps matter for business use. Voice with video or screen sharing still requires the legacy mode, and there is no API at launch — developers wanting to build on GPT-Live-1 join a waitlist, and Business, Enterprise and Edu workspaces are excluded for now. Brands hoping to wire the new conversational quality into their own products cannot yet.

OpenAI has also built voice-specific safeguards, a pointed move given the scrutiny of AI companions. The system can steer a conversation towards safer ground while the model is mid-sentence, surface support resources if topics such as self-harm arise, give age-appropriate responses to teenagers, and let parents disable voice for their children entirely. Voice conversations are automatically excluded from model training, with audio stored for 30 days and deletable, and the company was explicit that it is not building a companion product.

The competitive field is crowding at the same moment. Google’s Gemini Live has pursued the same natural-conversation ground, Apple and Amazon have rebuilt their assistants around better context handling, and OpenAI shipped GPT-Realtime-2.1 voice models to its developer API on 6 July — a sign the pipeline behind the consumer launch is already moving.

How marketing teams should respond

Marketing and CX teams do not need to wait for the API to act on this. The immediate work is testing: how a brand’s existing voice and chat experiences handle interruption, mid-sentence correction and escalation, because GPT-Live-1 has just moved the standard customers will measure them against. Disclosure and accessibility deserve the same scrutiny — a voice agent that sounds convincingly human raises the bar for making clear it is not one.

The strategic signal sits in OpenAI’s own framing. Atty Eleti, ChatGPT Voice’s product lead, said the company sees voice becoming “a kind of primary interface to computing”, including for long-running agentic work. If that holds, the brand interactions that today happen through screens — comparison, configuration, purchase — will increasingly happen through conversations a brand does not host and cannot fully script, which makes the accuracy of what assistants say about a company a marketing concern rather than a support one.

The API waitlist is open but undated, and OpenAI has not said which languages the translation feature formally supports. GPT-Live-1’s real test — third parties building customer experiences on it — has not started yet.