OpenAI has released a new family of real-time voice models, GPT-Live, designed to let people converse with artificial intelligence as they would with another person — talking, listening and interrupting without the awkward pauses that have long given away the machine.

The company said the models, GPT-Live-1 and a smaller GPT-Live-1 mini, can listen and speak at the same time, a capability that has separated human conversation from machine interaction since the first voice assistants appeared. OpenAI introduced the system on July 8 and detailed application-programming-interface access this week, framing it as the next step in moving AI out of text boxes and into natural speech.

The release extends a crowded race among leading laboratories to make voice the primary interface for assistants. Rivals including Anthropic and Google have pushed their own conversational systems, and the arrival of simultaneous listen-and-speak audio raises the stakes for customer service, education and accessibility, where a stilted back-and-forth has been a persistent barrier.

OpenAI said GPT-Live improves on earlier voice models in naturalness and can reason, translate across languages and transcribe speech in real time. The mini variant is aimed at lower-cost, higher-volume deployments, a sign the company expects voice to move from showcase to infrastructure rather than remaining a premium feature.

Voice also brings new risks. Synthetic speech that is indistinguishable from a human invites deepfakes, consent questions and concerns about always-on microphones. Regulators in several jurisdictions are weighing rules for cloned and generated voices, and while OpenAI has said safety tuning is built in, researchers note that real-world misuse is difficult to prevent completely.

Image source: i.ibb.co