OpenAI launches GPT-Live voice models for simultaneous ChatGPT conversation

OpenAI has launched GPT-Live, a new generation of voice models designed to enable full-duplex conversations. The update, which began rolling out globally on July 8, 2026, allows ChatGPT to listen and speak simultaneously, eliminating the turn-taking delays found in previous versions of the service. By processing bidirectional audio, the models can respond to interruptions and provide verbal cues like "mhmm" or "yeah" while a user is still speaking.

The rollout includes two distinct tiers: GPT-Live-1, which is available to paid subscribers, and GPT-Live-1 mini, which serves as the default voice mode for free users. These models are appearing across ChatGPT mobile apps and the web interface, with the company expecting full availability for all users within a few days of the initial release. The models utilize a continuous interaction framework to process audio input while generating speech, reducing the turn-based feel of previous assistants.

The architecture behind GPT-Live focuses on continuous interaction. According to OpenAI, the system can delegate complex tasks — such as web searches or intensive reasoning — to background models like GPT-5.5 while maintaining the flow of the ongoing voice conversation. When a background process concludes, the system weaves the information back into the active dialogue. Atty Eleti, product lead for ChatGPT voice, characterized this shift as a move toward voice becoming a primary interface for computing. Eleti stated:

"Over time, we think this will also unlock the ability to use voice as a kind of primary interface to computing, and to manage increasingly complex long-running agentic work."

In addition to conversational improvements, the update introduces real-time translation capabilities. The models are also capable of generating visual responses, such as graphics for weather reports or sports scores, in instances where visual information is more effective than spoken output. OpenAI noted that while the new mode is optimized for most spoken languages, some may exhibit non-native accents or variations in fluency.

The release of GPT-Live follows a period of safety-focused testing. OpenAI indicated that the new models include enhanced safeguards designed to provide age-appropriate responses and support for users discussing sensitive topics like self-harm. As part of its privacy policy, OpenAI stores audio clips for 30 days to maintain conversational context, though the company allows users to delete this data and automatically opts users out of AI training for voice mode interactions. Additionally, OpenAI has separate real-time voice models, GPT-Realtime-2 and GPT-Realtime-Whisper, available through its Realtime API, which are designed for developers and support methods such as WebRTC, WebSocket, and SIP.

This deployment occurs during a busy product week for the company. The rollout of the GPT-Live models is distinct from the scheduled release of the GPT-5.6 model series, which comprises the models Sol, Terra, and Luna. The GPT-5.6 series was delayed previously following a request from the Trump administration due to cybersecurity concerns. Following the completion of required safety testing, OpenAI announced it would proceed with the public release of the GPT-5.6 series on July 9, 2026.

OpenAI’s new voice capabilities enter a competitive market for low-latency AI assistants. Other providers, including Google and specialized startups, continue to develop similar full-duplex technologies. As the company continues to optimize its models, it is aiming to refine the voice experience for long-running, agentic tasks, and is currently maintaining a waitlist for developers interested in integrating GPT-Live technology through the API.