GPT-Live - A new-generation voice model launched by OpenAI

What is GPT-Live

GPT-Live is a new-generation voice model launched by OpenAI, using a full-duplex architecture to listen and speak simultaneously, making multiple decisions per second to speak, listen, interject, or call tools. The model decouples real-time interaction from deep tasks, delegating complex issues to the background GPT-5.5 for processing while the foreground maintains smooth conversation. GPT-Live has become the default voice mode of ChatGPT, offering two tiers, GPT-Live-1 (for paid users) and mini (for free users), covering iOS, Android, and web, and supporting visual cards such as weather and stocks.

Main features of GPT-Live

• Full-duplex continuous conversation: Supports listening and speaking simultaneously, no longer taking turns, and supports natural interruption, interjection, and real-time translation.
• Intelligent decision engine: Makes multiple judgments per second on whether to speak, listen, pause, interject, or call tools.
• Foreground-background task decoupling: Real-time interaction is handled by GPT-Live, while complex tasks such as search/reasoning are delegated to GPT-5.5 for background execution, keeping the foreground conversation continuously smooth.
• Visual responses: Supports rich visual cards for topics such as weather, stocks, and sports.
• Multiple reasoning tiers: Instant tier responds quickly, while Medium and High tiers enable GPT-5.5 Thinking deep reasoning.
• Background noise reduction: Focuses on the user's voice and reduces interference from ambient noise.
• Multilingual support: Optimized for commonly used languages; some languages may have a non-native accent.

Technical principles of GPT-Live

• Full-duplex continuous interaction architecture: GPT-Live uses a full-duplex architecture, where the model continuously processes audio input while generating voice output, no longer processing conversations in discrete turns. This enables it to make multiple interaction decisions per second, including speaking, continuing to listen, pausing to wait, naturally interjecting, or calling tools, achieving truly synchronized listening and speaking and solving the problem of interruptions caused by silence detection in traditional turn-based models.
• Foreground-background task decoupling mechanism: The system decouples real-time voice interaction from deep cognitive tasks: the foreground GPT-Live is dedicated to maintaining conversational fluency, and when encountering complex problems requiring search, multi-step reasoning, or Agent execution, it automatically delegates the task to the background GPT-5.5 for processing. After the background completes it, the result is seamlessly brought back into the conversation, and the foreground never interrupts the exchange, realizing a collaborative mode of “computing while chatting.”
• Real-time decision-making and tool calling: The model has a built-in high-frequency decision engine that judges interaction status in real time based on a continuous audio stream. It can remain silently waiting when the user pauses to think, insert feedback such as “mm-hmm” when supplementation is needed, and immediately call search or calculation capabilities when a tool need is identified. This second-level response mechanism makes the conversation rhythm closer to real human interaction rather than mechanical turn-taking.
• Native audio end-to-end processing: Unlike early cascaded architectures (STT→LLM→TTS) and turn-based end-to-end models, GPT-Live uses native audio streams as input and output, directly modeling the mapping between acoustic features and semantic content, avoiding information loss caused by text transcription. It also supports background noise reduction and sound source focusing, accurately identifying the user's voice within a continuous audio stream.

How to use GPT-Live

• Enter voice mode: Visit ChatGPT Appor the web version, and click the microphone icon at the bottom to enter voice conversation.
• Select reasoning tier: Paid users use GPT-Live-1 by default and can switch between three tiers: Instant, Medium, and High; free users use GPT-Live-1 mini by default.
• Natural conversation: Just start speaking. You can interrupt, pause, and interject, and the model responds in real time. When search or complex tasks are needed, the conversation continues in the foreground while GPT-5.5 is automatically invoked in the background to handle it.
• View visual cards: When asking about weather, stocks, sports events, and other content, the corresponding visual information cards are automatically displayed on the screen.
• API access (coming soon): Developers can visit the OpenAI official website https://openai.com/form/gpt-live-1-in-the-api/ to fill out the GPT-Live API registration form and wait for notification.

GPT-Live's core advantages

• Truly natural human-machine conversation: The model's full-duplex architecture eliminates the stiffness of turn-taking, supports interruptions and pauses for thought, and delivers an experience close to talking with a real person.
• Zero-interruption task handling: The foreground-background decoupled design lets complex tasks run in the background, so the foreground conversation never drops.
• Continuous evolution capability: The voice model and the reasoning model are decoupled, so future frontier model updates require no retraining of the voice system.
• Significantly better than the previous generation: It comprehensively surpasses the advanced voice mode in blind tests for conversation fluency, pleasantness, interruption handling, and more.
• Visual enhancement: Voice interaction is combined with graphic cards, making information presentation more intuitive.

GPT-Live's project address

• Project official website：https://openai.com/zh-Hans-CN/index/introducing-gpt-live/

Comparison of GPT-Live with similar competitors

GPT-Live's application scenarios