OpenAI Launches GPT-Live: Full-Duplex Voice Models That Can Listen and Talk at the Same Time

OpenAI on Tuesday unveiled GPT-Live, a new family of voice models designed to make AI conversation feel less like querying a database and more like talking to a person. The launch, which arrived with little advance warning, replaces the company’s existing Advanced Voice Mode in ChatGPT with a fundamentally different architecture — and sets a new competitive bar for real-time AI interaction.

A Different Kind of Listening

The defining technical leap in GPT-Live is full-duplex audio processing. Previous voice AI pipelines — including OpenAI’s own — worked sequentially: speech would be converted to text, fed to a language model, and then rendered back to audio. Each step added latency, and the system could only either listen or speak at any moment, not both.

GPT-Live tears down that pipeline. The new models process incoming speech and generate outgoing audio simultaneously, making interaction decisions many times per second: whether to speak, continue listening, pause, inject a brief affirmation, or hand off a query to a downstream tool. The result is a system that can catch a user mid-sentence, respond to a correction without waiting for silence, and sustain conversation rhythms that previously required a human on the other end.

“Over time, we think this will also unlock the ability to use voice as a primary interface to computing,” said Atty Eleti, product lead for ChatGPT Voice. That framing — voice as interface, not just feature — reveals OpenAI’s longer-term ambitions.

Two Models, Tiered Access

OpenAI released two versions simultaneously. GPT-Live-1 mini is now the default voice experience for all ChatGPT users, replacing the previous Advanced Voice Mode that had been in place since mid-2024. GPT-Live-1, a larger and more capable model, is available to paid subscribers on Plus, Pro, and Team plans.

Both models are built on top of GPT-5.5, OpenAI’s current frontier model. For questions that require web search, extended reasoning, or agentic tool use, GPT-Live can delegate to GPT-5.5 in the background and weave the results back into the spoken response — preserving conversation flow without forcing users into a text-based interface for complex queries.

The system supports extended conversation sessions of 30 to 40 minutes — a notable improvement over earlier voice modes that often timed out at 10 minutes or less. Visual information, such as charts or reference images, can also be surfaced alongside voice responses.

Why This Matters Now

The timing of the GPT-Live launch is not accidental. OpenAI is racing to establish a dominant position in voice AI before a cluttered field of competitors — including Google’s Gemini Live, Apple Intelligence’s upgraded Siri, and a growing roster of specialized voice startups — can gain meaningful consumer traction.

Voice is increasingly seen as the next major interface shift in consumer computing, particularly for mobile. OpenAI’s integration with Apple CarPlay, confirmed in the same week as the launch, signals that the company is pushing GPT-Live beyond smartphones into contexts where screen interaction is impractical or unsafe. CarPlay integration means GPT-Live could become the default conversational layer for hundreds of millions of drivers worldwide.

The full-duplex architecture also opens doors to real-time translation, a feature OpenAI confirmed is in development. Early testing suggests the system can interpret and relay conversation across languages with acceptable latency, though OpenAI has not yet made this feature broadly available.

Teenage Guardrails and Safety Infrastructure

OpenAI also shipped age-appropriate safeguards in the same release. For verified teen accounts, GPT-Live applies stricter content filters and more conservative behavioral defaults than for adult users. The company did not provide specifics on how age verification is enforced, a detail that is likely to draw scrutiny from regulators given the difficulty of reliable age verification at scale.

The teen safety layer arrives as AI voice products face increasing regulatory pressure. The EU’s AI Act, now in enforcement across all 27 member states, requires that AI systems likely to interact with children carry enhanced safety measures and disclosure requirements. Illinois signed its own AI safety law just two days before the GPT-Live launch, and multiple U.S. states are considering comparable measures.

A Precursor to GPT-5.6

The GPT-Live launch is widely understood to be a staging move. OpenAI is expected to release GPT-5.6 Sol, its next major text model, in the weeks ahead. The voice-first rollout appears designed to warm up the user base for that release, establishing a cohesive product story: a frontier model paired with a dedicated, highly optimized voice layer.

Analysts tracking OpenAI’s enterprise business have noted that the ability to run extended, context-rich voice sessions is particularly valuable in customer service, sales, and healthcare. Several enterprise clients are reported to be piloting GPT-Live through the Realtime API, which exposes the full-duplex capabilities to developers building custom voice agents.

The Competitive Picture

Google’s Gemini Live, which launched broadly in early 2026, has accumulated millions of monthly active users and benefits from deep integration across Android devices. Apple Intelligence’s enhanced Siri, due to ship fully in iOS 26.6, is likely to compete directly for the same conversational voice niche. Meta has also previewed voice capabilities for its Ray-Ban smart glasses lineup, targeting ambient, always-on scenarios that neither OpenAI nor Google has directly addressed.

The difference, OpenAI argues, is the underlying model quality and the ecosystem depth. GPT-Live-1’s connection to GPT-5.5’s tool-use infrastructure means it can take actions — drafting emails, searching the web, setting reminders — in the middle of spoken conversation, not just respond to queries. For users already embedded in the OpenAI ecosystem, that integration represents a meaningful advantage.

Whether GPT-Live is enough to lock in voice-first users before competitors can match it remains an open question. What is clear is that the race to become the dominant conversational layer in everyday computing has entered a new, faster phase — and OpenAI intends to lead it.