OpenAI makes ChatGPT better at banter

AI and ML

With GPT-Live, talking, listening, and formulating answers all happen at once

OpenAI has released a new voice model that can produce human-sounding speech, or scour the web in response to spoken queries.

GPT-Live, according to the company, makes chatbot banter feel more like a real conversation, something of a bold move for a company battling multiple lawsuits alleging mental health harms because people took ChatGPT too seriously.

"During conversations, GPT‑Live can show it’s paying attention with phrases like 'mhmm' or 'yeah', engage in quick back-and-forth, or just stay quiet when you need a moment to think," the company said in a blog post. "The result is a voice experience that is refreshingly easy to talk to."

The company has published a video demonstrating this full duplex experience. It features three women of an age seldom seen at companies like OpenAI but often impacted by the kinds of scams AI technology enables.

OpenAI insists that it has expanded its safety testing regime to better assess native audio interactions. And it has published a system card to document its approach.

While OpenAI notes that it has policies and protections against voice cloning and impersonation, the company has not disavowed replicating a competing product from former CTO Mira Murati.

Murati's company Thinking Machines in May talked up "interaction models" and how they can speak, listen, and search the web at the same time. Two months later, OpenAI has a similar offering.

If Apple were involved, we'd say Thinking Machines had been "Sherlocked," a term from a time when copying a startup's product stirred indignation. We'd suggest "Altmanned" as an alternative if it weren't for the global shrug of indifference to frontier model companies capturing the world's intellectual output, laundering it, and reselling it.

GPT-Live will delegate queries that require web search to a background model (GPT-5.5 presently) that processes the request while maintaining conversational flow with the user. The company's hope is that this will allow voice interaction to drive more complicated, lengthy agentic workflows – which tend to inflate token usage and billing.

Whether an original idea, a parallel innovation, or a sincerely flattering imitation, GPT-Live's full-duplex implementation represents an improvement in model architecture. "Instead of processing a sequence of separate messages, GPT‑Live continuously processes input while generating output," OpenAI explains. "The model can therefore make interaction decisions many times per second: whether to speak, continue listening, pause, interrupt, or invoke a tool."

It will be interesting to see whether security researchers find that this approach, of continuously processing input, allows for novel attack opportunities.

ChatGPT users can invoke GPT-Live by tapping the "Voice" button. OpenAI contends this experience will result in more natural conversations, better answers, improved listening, and visual feedback.

GPT-Live will appear in the iOS and Android ChatGPT apps, and on the web. A more capable version, GPT-Live-1, is the default for ChatGPT Voice for Go, Plus, and Pro users. Free-tier customers have to settle for GPT-Live-1 mini.

GPT-Live comes with a caveat – it has been optimized for popular languages and may not work all that well for "certain languages" yet. But now that the model has been made more responsive, any missteps should be noticeable a few milliseconds sooner. ®