OpenAI Launches GPT-Live Real-Time Voice Feature, Backed by GPT-5.5 for Natural Conversations

OpenAI has officially released GPT-Live, a next-generation voice model based on a full-duplex architecture designed to enhance the naturalness and intelligence of ChatGPT voice interactions. The model can simultaneously listen and speak, allowing users to interrupt at any time, pause to think, and maintain conversational flow with brief responses like "uh-huh" or "okay." It can delegate tasks to advanced models like GPT-5.5 in the background to handle complex needs such as search and reasoning, achieving a separation architecture where the frontend chats while the backend works.

Architecture Innovation

• Full-Duplex Continuous Interaction: GPT-Live continuously processes input while generating output, deciding multiple times per second whether to speak, listen, or interrupt, solving the interruption problem caused by silence detection in previous turn-based models.
• Deep Task Delegation: Decouples real-time conversation from complex tasks. The frontend GPT-Live handles low-latency interaction, while the backend GPT-5.5 (or subsequent models) handles tasks like search and reasoning, with results naturally integrated into the conversation. Users can choose between Instant (fast), Medium, or High (deep thinking) modes.

Key Improvements

• More Natural Conversations: Supports real-time interruption, pauses for thinking, and focuses on the user's voice even with background noise. OpenAI has refined 9 default voice tones.
• Smarter Responses: With GPT-5.5 integration, reasoning and search capabilities are significantly enhanced. In official demos, the model can perform real-time translation, weather queries, and interview reviews.
• Visual Replies: Voice conversations can display visual cards for weather, stocks, etc., supporting web search, memory, images, and file uploads.

Evaluation and Data

• OpenAI established a new human evaluation system, with GPT-Live significantly outperforming Advanced Voice Mode in pleasantness and fluency. Conversation fluency scores improved from 3.80 to 4.96 (out of 7).
• GPT-Live outperforms its predecessor on benchmarks such as GPQA (scientific reasoning), BrowseComp (web search), and τ³-Voice Telecom (customer service tasks).
• Over 150 million users use ChatGPT voice features weekly.

Release and Limitations

• Versions: GPT-Live-1 (default, for Go/Plus/Pro users) and GPT-Live-1 mini (for Free users).
• Platforms: iOS, Android, and ChatGPT.com.
• Limitations: Does not support video or screen sharing at launch, not available for B2B, and does not support desktop app, Codex, or custom GPTs. API pricing has not been announced.

Industry Comparison

• ByteDance Doubao: Full-duplex model Seeduplex, supports screen sharing and video, but no clear backend strong model delegation.
• MiniMax: Speech 2.8 focuses on TTS and voice cloning, suitable for content production.
• Shenbian Intelligence: MiniCPM-o 4.5 full-modal full-duplex, open-source and free.
• Google: Gemini 3.1 Flash Live Preview low-cost multimodal, approximately $0.036 per minute.
• ElevenLabs: Enterprise voice agent, $0.08 per minute.

GPT-Live pushes voice interaction to a more natural level through architectural innovation, but product details and pricing still need refinement.