OpenAI launched GPT-Live on 8 July 2026, a pair of full-duplex voice models (GPT-Live-1 and GPT-Live-1 mini) that can listen and talk at the same time inside ChatGPT. The release replaces Advanced Voice Mode as the default and marks a real architectural shift: instead of taking turns, the model holds a conversation the way a person does, with overlapping speech, backchannels, and pauses.
TL;DR
- GPT-Live launched 8 July 2026 with two models: GPT-Live-1 (Go, Plus, Pro) and GPT-Live-1 mini (Free default).
- Full-duplex means simultaneous listening and speaking, with backchannels and clean interruptions.
- Backend at launch is GPT-5.5, which handles search, reasoning, and agentic tasks in the background.
- Around 150 million people already use ChatGPT voice; OpenAI frames voice as a candidate for the primary computing interface.
- Legacy Standard and Advanced Voice Mode remain; video and screen sharing are planned for later.
What are the OpenAI GPT-Live voice models?
GPT-Live is a new family of voice-native models inside ChatGPT on iOS, Android, and ChatGPT.com. There are two variants at launch. GPT-Live-1 is available to paid users on the Go, Plus, and Pro tiers. GPT-Live-1 mini is the default voice for Free users and replaces the previous Advanced Voice Mode. Both are audio-native rather than text-to-speech wrappers around a chat model, which allows them to overlap speech and silence in the same stream.
Behind the voice layer sits GPT-5.5, which the voice model calls for web search, longer reasoning, and agentic actions. You can ask a rambling multi-part question, the voice model keeps the conversation open with short verbal cues, and the heavier work runs in the background before the answer comes back.
How does full-duplex voice actually work?
Traditional voice assistants are half-duplex: you speak, they wait, they speak, you wait. Full-duplex means input and output audio streams run in parallel. The model can hear you mid-sentence and choose whether to interject, wait, or answer once you finish. That opens three behaviours the older Advanced Voice Mode struggled with:
- Active listening. The model produces short backchannels such as "mhmm" or "yeah" while you are still talking, which signals attention without cutting you off.
- Long silences. It can stay quiet for extended stretches while you think aloud or dictate, then respond in context rather than assuming the turn has ended.
- Clean interruption. If you cut in, the model stops immediately and adjusts, instead of finishing its sentence over you.
ChatGPT Voice product lead Atty Eleti has said he has run 30 to 40 minute conversations with GPT-Live-1 without the interaction feeling like a series of prompts. That length is the point: full-duplex is less about faster replies and more about sustaining a single thread of talk.
What can GPT-Live do that Advanced Voice Mode could not?
The headline behavioural changes are conversational, but there are measurable ones too. GPT-Live-1 outperforms Advanced Voice Mode on GPQA (expert-level scientific reasoning) and BrowseComp (agentic web search), because the voice layer can hand off to GPT-5.5 mid-conversation instead of blocking on a single monolithic response.
Practical differences you will notice:
- Delegated tasks. Ask for a weather forecast, a stock quote, or a live sports score, and the model can display a visual card in the ChatGPT app alongside the spoken answer.
- Background work. While GPT-5.5 handles a longer search or reasoning step, the voice model keeps the channel open with acknowledgements rather than dead air.
- More natural pacing. Because the model is not forced to finish speaking before it can hear you, it can start a sentence, sense hesitation, and revise.
For a broader view of where the underlying reasoning model sits in OpenAI's lineup, see our analysis of the GPT-6 delay and GPT-5.5 release, the GPT-5.6 Sol launch guide, and the release-gate strategy tied to government approval.
Who gets GPT-Live and how do you use it?
Access is tier-based inside ChatGPT:
- Free users: GPT-Live-1 mini is now the default voice experience, replacing Advanced Voice Mode.
- Go, Plus, and Pro users: GPT-Live-1 is available and set as the default for voice sessions.
- Everyone: The legacy Standard and Advanced Voice Mode remain selectable for users who prefer the older behaviour or need it for a specific workflow.
To start a session, open the ChatGPT app on iOS or Android, or ChatGPT.com in a supported browser, and tap the voice icon. Voice with video and screen sharing is on the roadmap but not part of the 8 July release.
What are the real limitations of GPT-Live?
Two things are worth being honest about. First, non-English support is uneven. OpenAI's own Hindi demonstration showed a heavy American accent and delivery that did not sound like a native speaker. If your use case involves customer-facing calls in Hindi, Arabic, or other languages outside the model's strongest coverage, test before rolling anything out.
Second, full-duplex does not remove hallucinations. When the voice model delegates to GPT-5.5 for a search or a factual claim, it inherits whatever mistakes the reasoning layer makes. The conversational polish can make errors feel more authoritative, not less.
OpenAI has expanded testing to include audio-native evaluations, built in safeguards against voice impersonation, and added teen-specific protections. The full breakdown is in the GPT-Live system card.
Why does full-duplex voice matter for AI agents?
The strategic bet is that voice becomes the primary interface for agentic workflows, not just consumer chat. If an AI agent is doing work on your behalf, a natural back-and-forth conversation is a better control loop than typing prompts. You can supervise, correct, and delegate in the same rhythm you would with a colleague on a call.
That framing lines up with the rest of the industry. Apple has updated Siri, Amazon has updated Alexa, and startups such as Sesame (founded by former Oculus leadership) and Monogram (which raised a $40M seed) are building around always-on voice. For more on the wider agent shift, see our pieces on Claude's always-on agent era, AI agents as a 2026 business capability, and autonomous agent loops. For small teams, our guides on ChatGPT for small business and choosing an AI chatbot in 2026 cover adjacent decisions.
FAQ
Q: When did OpenAI GPT-Live launch? A: 8 July 2026, inside ChatGPT on iOS, Android, and ChatGPT.com.
Q: What is the difference between GPT-Live-1 and GPT-Live-1 mini? A: GPT-Live-1 is the higher-capability model for Go, Plus, and Pro subscribers. GPT-Live-1 mini is the default for Free users.
Q: Does GPT-Live replace Advanced Voice Mode? A: By default, yes. Legacy Standard and Advanced Voice Modes remain available if you prefer them.
Q: What model powers the reasoning behind GPT-Live? A: GPT-5.5 handles search, reasoning, and agentic actions while the voice layer manages the conversation.
Q: Can GPT-Live handle video or screen sharing? A: Not yet. Both are planned for a future release.
Q: How well does GPT-Live work in languages other than English? A: Uneven at launch. OpenAI's own Hindi demo showed a heavy American accent and unnatural speech. Test before shipping to non-English users.
Discussion
0 comments