OpenAI just made ChatGPT’s voice feel way more like a conversation than a sequence of prompts. GPT-Live is a new voice model that can speak in human-sounding audio, listen while it’s generating, and handle spoken questions with quick back-and-forth, or even pause and stay quiet when the user needs a moment to think. In OpenAI’s own description, during conversations GPT-Live can show it’s paying attention with phrases like “mhmm” or “yeah.” And because it is “full duplex,” it continuously processes input while generating output, letting it make interaction decisions many times per second about whether to speak, continue listening, pause, interrupt, or invoke a tool.

This matters right now because OpenAI has not been operating in a calm public-health vacuum. The company is “battling multiple lawsuits alleging mental health harms because people took ChatGPT too seriously,” and a more human-sounding, more responsive voice experience can change how people perceive and engage with the system. OpenAI’s pitch is that GPT-Live makes voice interaction “refreshingly easy to talk to,” and it has published a video demonstrating the full duplex experience. The demo also features three women of an age seldom seen at companies like OpenAI, which is relevant because the same kind of voice-driven AI can be used in scams, and the source makes that connection explicit.

Under the hood, GPT-Live routes work that requires web search to a background model (GPT-5.5 presently) while maintaining the conversational flow with the user. That design choice is a big deal for product and cost governance. OpenAI’s hope is that voice interaction can drive more complicated, lengthy agentic workflows, which tend to inflate token usage and billing. In other words, this is not just “sound better.” It is architected to support more of the end-to-end work that makes agents expensive: planning, tool use, iterative reasoning, and longer multi-step outputs, all happening while the user experiences a smooth, turn-by-turn conversation.

The company frames the architectural upgrade as a shift away from processing a sequence of separate messages. Instead of handling discrete chunks and waiting for each one to complete, OpenAI explains that GPT-Live “continuously processes input while generating output.” That continuous processing is what enables the many-times-per-second interaction decisions. It is also what can shake up the security conversation, because continuous input processing is not merely a UX upgrade. The source points out that it will be interesting to see whether security researchers find novel attack opportunities stemming from this approach of continuously processing input, which could create different timing, control, or interruption dynamics than traditional message-based flows.

OpenAI is not ignoring safety. It says it has “expanded [its] safety testing regime to better assess native audio interactions,” and it has published a system card to document its approach. The source also notes OpenAI’s policies and protections against voice cloning and impersonation, which are essential in a world where realistic speech can be abused. But there is another layer of friction: the company has not disavowed replicating a competing product from former CTO Mira Murati. The source ties this directly to Thinking Machines, which in May talked up “interaction models” that can speak, listen, and search the web at the same time. Two months later, OpenAI has a similar offering.

For executives, this is where product strategy, competitive positioning, and legal risk collide. The source even references “Sherlocked,” and then coins “Altmanned” as an alternative, pointing to a broader pattern in frontier model ecosystems where startups’ intellectual output can be captured, repackaged, and resold. Whether or not you buy the slang, the practical implication is clear: when model architectures and interaction capabilities converge quickly across companies, boards and leadership teams get pressure from multiple directions. You get scrutiny from customers and regulators, you get diligence questions from investors, and you get reputational risk from disputes that may not resolve neatly.

GPT-Live will be available inside ChatGPT: users can invoke it by tapping the “Voice” button, and it will appear in the iOS and Android ChatGPT apps and on the web. A more capable version, GPT-Live-1, will be the default for ChatGPT Voice for Go, Plus, and Pro users. Free-tier customers get GPT-Live-1 mini. OpenAI also includes a caveat: GPT-Live has been optimized for popular languages and “may not work all that well for ‘certain languages’ yet.” Translation: even as responsiveness improves, missteps may show up sooner, which matters if you are rolling out voice experiences in markets with less coverage.

Second-order stakes for decision-makers: voice is not a peripheral feature anymore. It is a new interface layer for agentic workflows, with expanded tool use and web search delegation happening behind the scenes while the user experiences continuous, natural interaction. If that increases capability, it can also increase liability, because “more human” can mean “more trusted,” and “more trusted” can intensify the harms that lawsuits allege. GPT-Live’s full-duplex implementation therefore lands as both an architecture upgrade and a governance stress test, setting expectations for everyone building, regulating, or purchasing conversational AI.