OpenAI Launches GPT-Live, a Full-Duplex Voice That Listens and Talks at Once

New voice models let ChatGPT hear and speak simultaneously, hand hard questions to GPT-5.5 in the background, and keep the conversation going — replacing Advanced Voice Mode as the default for 150 million-plus voice users.

Jul 9, 2026Source: TechCrunch

OpenAI Launches GPT-Live, a Full-Duplex Voice That Listens and Talks at Once

Key Points

1.OpenAI on July 8 introduced GPT-Live — models GPT-Live-1 (for paid Go, Plus and Pro) and GPT-Live-1 mini (for Free) — a full-duplex system that listens and speaks at the same time, rolling out globally on iOS, Android and the web over the coming days.
2.For complex requests needing web search, deeper reasoning or agentic work, GPT-Live delegates to GPT-5.5 asynchronously and keeps talking while the answer computes; it remasters ChatGPT's nine voices and adds Instant, Medium and High reasoning settings.
3.OpenAI published a GPT-Live System Card detailing voice-specific safeguards — real-time input/output checks, spoken safety messages and red-teaming for impersonation, self-harm and emotional reliance — as some early users called the model's 'mhmm' filler acknowledgments annoying.

OpenAI is rebuilding how ChatGPT talks. On July 8 the company introduced GPT-Live, a pair of voice models — GPT-Live-1 for paid Go, Plus and Pro users and GPT-Live-1 mini for the Free tier — that listen and speak at the same time instead of taking turns. It replaces Advanced Voice Mode as the default and is rolling out globally on iOS, Android and ChatGPT.com "over the next few days."

Full duplex, explained

The core shift is architectural: GPT-Live continuously processes what it hears while generating speech, making decisions "many times per second" about whether to talk, keep listening, pause, interrupt or call a tool. It offers natural acknowledgments — "mhmm," "got it," "yeah" — handles interruptions, and can be asked to slow down. OpenAI remastered ChatGPT's nine voices for the system and added three reasoning levels: Instant, Medium and High.

The delegation trick

GPT-Live is tuned for fast, natural conversation, so when a request demands web search, deeper reasoning or agentic work, it hands the task to GPT-5.5 in the background and keeps chatting while the computation runs, then folds the result back in. It also surfaces visual cards for things like weather, stocks and sports. What it doesn't do at launch: voice-plus-video, screen sharing, desktop or Codex, and Business, Enterprise and Edu tiers wait for a later phase.

Safety for a new modality

OpenAI published a dedicated GPT-Live System Card. Inputs and outputs are checked as the conversation unfolds; when something risky surfaces, the system can steer or interrupt a response, play a spoken safety message, offer text-based support resources, or end the call in higher-risk cases. Red-teaming spanned impersonation, speaker identification, child-coded voice, self-harm, emotional reliance and scams. OpenAI says the models are "not positioned as an AI companion."

Reception

More than 150 million people already use ChatGPT's voice features, and OpenAI's own evals show GPT-Live "strongly preferred" over Advanced Voice Mode in matched conversations, at roughly a 75% preference rate. "Voice can be the future interface to all kinds of work," said ChatGPT Voice lead Atty Eleti. Not everyone is sold: some early users griped that the constant "mhmm" filler is verbose and grating. An API version is listed as "coming soon," with no pricing yet.

Why It Matters

Full-duplex voice plus background delegation is OpenAI's bid to make talking — not typing — the default way people use AI, and to defend that surface before Google's Gemini, Amazon's Alexa revamp and Apple's Siri get there. The delegation pattern also quietly hides latency, the biggest thing that has made voice assistants feel robotic; if it works, it resets the bar for every rival's voice product.

Filed underOpenAI AI Tools #AI safety #ChatGPT #GPT-Live #voice AI