OpenAI GPT-Live Release Notes

GPT-Live introduces full-duplex voice interaction

OpenAI has launched GPT-Live, a new generation of voice models designed to make AI conversations feel fluid and natural. Unlike previous iterations, GPT-Live utilizes a full-duplex architecture, allowing the model to listen and speak simultaneously. This enables the AI to provide active listening cues (such as "mhmm" or "yeah"), handle quick back-and-forth exchanges, and remain silent when a user is thinking, rather than interrupting based on detected silence.

Architectural Shift: From Cascaded to Full-Duplex

GPT-Live represents a significant departure from two previous voice AI paradigms:

1. Cascaded Voice Systems: These relied on a chain of three separate models: speech-to-text (transcription), a large language model (response generation), and text-to-speech (audio conversion). This process often resulted in slow, stilted responses and loss of information between stages.
2. Turn-based Voice Models: While these processed audio within a single model to reduce latency, they still operated in discrete turns. The model had to wait for the user to stop speaking before responding, often leading to rigid interactions or accidental interruptions caused by background noise.

GPT-Live solves these issues by continuously processing input while generating output. The model makes interaction decisions—such as whether to speak, pause, or invoke a tool—multiple times per second.

Intelligent Delegation to GPT-5.5

To maintain a natural conversational flow without sacrificing intelligence, GPT-Live decouples interaction from deep reasoning. When a request requires web search, complex reasoning, or agentic capabilities, GPT-Live delegates the task to a frontier model—specifically GPT-5.5 at launch—in the background.

This delegation allows the AI to keep talking and maintaining the flow of the conversation while the heavier reasoning process occurs. Users can choose their preferred level of reasoning effort:

• Instant: For fast, immediate responses.
• Medium and High: For tasks requiring more thinking time and deeper reasoning.

Performance and Safety Benchmarks

In head-to-head human evaluations, GPT-Live-1 and GPT-Live-1 mini were strongly preferred over Advanced Voice Mode across metrics including turn-taking, conversational flow, and overall naturalness. Technical benchmarks also showed gains:

• GPQA: Substantial outperformance of Advanced Voice Mode in expert-level scientific reasoning.
• BrowseComp: Strong gains in agentic web search and locating difficult information.
• τ³-Voice Telecom: Superior performance in realistic, multi-turn telecom support tasks.

Safety and Safeguards

GPT-Live includes dedicated safety training and audio-native evaluations focusing on risks such as self-harm, violence, and sexual content. Key safety features include:

• Real-time steering: The system can steer the model toward safer responses or end the conversation if unsafe output is detected while the model is speaking.
• Teen protections: Age-appropriate behavior is trained into the model, and parental controls allow parents to manage access and receive notifications for high-risk situations.
• Impersonation prevention: The model uses predefined voices and includes safeguards to prevent the imitation of real people's voices.

Availability and Features

GPT-Live is rolling out globally across iOS, Android, and ChatGPT.com.

• GPT-Live-1: Default for Plus, Pro, and Go users.
• GPT-Live-1 mini: Default for Free users.

New features include rich visual cards for weather, stocks, and sports, as well as continued support for search, memory, images, and file uploads. While voice with video and screen sharing is not supported at launch, OpenAI plans to introduce these capabilities in the future.

Community Insights and User Feedback

Early users and community members have highlighted both the potential and the limitations of the new system:

"The best feature is that it can delegate questions out to GPT-5.5 in the background, so you're no longer restricted to a voice model that's several years behind the frontier."

Key observations from the community include:

• Utility for Accessibility: Blind users have noted that combining full-duplex voice with future video/glasses integration could revolutionize navigation and assistance.
• Latency and Interruption: Some users observed that the model occasionally interrupts too quickly or fails to distinguish between a direct command and an aside (e.g., translating "tell him that..." verbatim).
• Reasoning Disparity: One user reported that voice responses, even when invoking GPT-5.5, felt "hand wavy" and lacked the detail found in direct text chat for complex technical questions.
• Behavioral Quirks: Reports include the model acting "aggressively" to moderate tone or interrupting with laughter, though some users noted these behaviors may have been clamped down in updates.