OpenAI has announced GPT-Live, an advanced voice model featuring full-duplex architecture that allows simultaneous listening and speaking, enabling natural conversations and live translation. It can delegate complex tasks to a frontier model when needed. The model is rolling out to ChatGPT users on iOS, Android, and web over the next few days.
SynthePulse Insight · AI deep reading
Version 1 · 9 sources
OpenAI releases the GPT-Live series of voice models, featuring a full-duplex architecture that enables simultaneous listening and speaking, and by default replaces the Advanced Voice Mode in ChatGPT. The new models can call on text models like GPT-5.5 for complex tasks, but a live demo of Hindi translation still sounded stilted.
The core of the GPT-Live series is a full-duplex architecture that can listen and speak at the same time, supporting natural interruptions and real-time translation. The previous Advanced Voice Mode relied on a pipeline: speech-to-text → language model → text-to-speech, while the new model processes end-to-end directly. OpenAI stated on X that the full-duplex architecture allows the model to conduct more natural back-and-forth conversations, maintain better timing, and even perform real-time translation.
OpenAI launched two models: GPT-Live-1 and GPT-Live-1 mini. GPT-Live-1 mini by default replaces the Advanced Voice Mode in ChatGPT, and is already available to all Go, Plus, and Pro users, with free users gradually receiving access. Paid users can access the larger GPT-Live-1 model. The new models are accessed via the voice button on iOS, Android, and the web version of ChatGPT, with an API coming soon.
GPT-Live can send queries requiring search, reasoning, or agent capabilities to the latest text models like GPT-5.5, wait for results in the background, and bring them back into the conversation. OpenAI stated in a briefing that this solves the old model's problems of interrupting users and lacking intelligence. Product lead Atty Eleti said he has had 30- to 40-minute walking conversations with the voice feature and believes voice could become the primary computing interface for complex work.
TechCrunch reported that during a live demo of real-time translation, Hindi translation had a heavy American accent, unnatural intonation, and sounded somewhat bookish. OpenAI said the new models are optimized for 'most languages' but did not specify which. This shows that despite the advanced architecture, non-English language quality still needs improvement.
OpenAI emphasizes that GPT-Live is not an AI companion but aims to enable more natural task interactions. The new models include built-in safety measures, such as providing age-appropriate responses for teens and offering resources when conversations involve self-harm. The company also noted that over 150 million people already use ChatGPT's voice and dictation features, indicating widespread demand for voice interaction.
Key information comes from OpenAI's official blog, X account, and TechCrunch reports. The Hindi translation issue in the demo was observed firsthand by a TechCrunch reporter, making it a primary report. OpenAI's statements about model capabilities and safety measures are from official sources but lack independent verification.
GPT-Live marks a shift in voice AI from pipeline architecture to full-duplex architecture, making progress in natural interaction and intelligent integration, but non-English language quality remains a weak point. OpenAI clearly positions it as a task tool rather than an emotional companion and emphasizes safety design.
Primary report
Primary source
Also covered by 5 sources