OpenAI replaces Advanced Voice Mode with full-duplex GPT-Live

OpenAI has launched GPT-Live, a new generation of voice models designed to make conversations with ChatGPT feel more human. The company announces in an official blog post that GPT-Live replaces Advanced Voice Mode and rolls out globally on iOS, Android, and ChatGPT.com.

The core innovation is a full-duplex architecture. Unlike previous systems, which waited for users to finish speaking before responding, GPT-Live processes incoming audio and generates output simultaneously. This lets the model decide, many times per second, whether to speak, stay quiet, or interrupt, according to OpenAI. It can insert small acknowledgments like “mhmm” or “yeah” while a user is still talking, without derailing the conversation.

Complex tasks run in the background

A second structural change separates conversation from reasoning. When a question requires web search or deeper thinking, GPT-Live delegates the task to OpenAI’s frontier model, currently GPT-5.5, and keeps talking while that model works in the background. Michael Nuñez writes for VentureBeat that this modular design allows OpenAI to upgrade the model’s intelligence without retraining the voice system itself.

Sabrina Ortiz writes for The Deep View that this combination was demonstrated live: GPT-Live translated a presentation into Hindi in real time while the speaker was still talking, without the awkward pause that plagued earlier versions.

Two versions, tiered rollout

• GPT-Live-1 becomes the default voice model for Go, Plus, and Pro subscribers.
• GPT-Live-1 mini becomes the default for free users.
• Both use GPT-5.5 Instant in the background, while higher reasoning tiers (Medium, High) tap into GPT-5.5 Thinking.
• API access for developers is planned but not yet available.

OpenAI reports that GPT-Live outperforms Advanced Voice Mode in human preference tests as well as on benchmarks such as GPQA (scientific reasoning) and BrowseComp (agentic web search).

Safety and lingering questions

OpenAI says it expanded audio-specific safety testing, covering categories like self-harm, emotional reliance, and sexual content. Nuñez notes that on synthetic adversarial tests, safety scores rose sharply compared to Advanced Voice Mode, though production-prompt evaluations showed a slight, statistically insignificant regression on emotional reliance. OpenAI is also rolling out long-term monitoring of emotional reliance and says the system uses only predefined voices to prevent impersonation, a response to the 2024 controversy over a voice resembling Scarlett Johansson’s.

Ortiz cautions that a more human-sounding assistant raises its own risks: people may trust it more readily, even though it can still hallucinate like any other AI model.

GPT-Live does not yet support video or screen sharing, and language fluency varies, OpenAI acknowledges. Nuñez points out that rivals including Google’s Gemini Live, ByteDance’s Seeduplex, and Nvidia’s PersonaPlex have already released full-duplex voice systems, meaning OpenAI’s upgrade closes a gap rather than opening a lasting lead.

Sources

• Introducing GPT-Live – OpenAI
• How OpenAI’s voice assistant got more natural – The Deep View
• OpenAI launches GPT-Live, a full-duplex voice upgrade that lets ChatGPT talk more like a person – VentureBeat