GPT-Live Ships Full-Duplex Voice on ChatGPT
- What happened
- OpenAI launched GPT-Live-1 (paid) and GPT-Live-1 mini (free) on July 8 — full-duplex voice models that listen and speak simultaneously, replacing Advanced Voice Mode across all ChatGPT tiers globally.
- Why it matters
- This is the third generation of ChatGPT voice in two years, and it changes the interaction model fundamentally: from turn-based Q&A to continuous conversation with background reasoning via GPT-5.5. With 150M weekly voice users, it is the largest voice AI deployment ever.
- What to do
- Paid users get GPT-Live-1 now. Free users get GPT-Live-1 mini. Try it — the continuous listening and natural interruption handling is the difference between a voice interface and a voice conversation. Developers: sign up for API access.
OpenAI shipped GPT-Live on July 8 — a pair of full-duplex voice models that replace Advanced Voice Mode as ChatGPT's conversational engine. GPT-Live-1 (conditional) serves paid users; GPT-Live-1 mini (conditional) rolls out to the free tier. Both listen and speak simultaneously, ending the two-year turn-based era of AI voice (OpenAI, 2026(opens in new tab)).
What happened
The old Advanced Voice Mode (launched in 2024) collapsed the speech-to-text → LLM → text-to-speech pipeline into a single model, but conversations were still turn-based. The model waited for you to stop speaking. GPT-Live is different: a full-duplex architecture that processes input while generating output — it can listen and speak at the same time.
This changes the interaction model in three ways:
- Natural interruption. You can jump in mid-response the way you would with a person. The model handles turn-taking dynamically instead of waiting for endpoint detection.
- Background reasoning. GPT-Live delegates complex tasks — web search, deep reasoning, agentic work — to GPT-5.5 in the background while keeping the conversation flowing. You get a verbal "let me think about that" instead of dead air.
- Real-time translation. The full-duplex pipeline can translate simultaneously rather than waiting for the speaker to finish a sentence.
OpenAI described the architecture as a full-duplex model that "continuously processes input while generating output" (OpenAI, 2026(opens in new tab)).
The rollout comes with new safety infrastructure. OpenAI built audio-native safety evaluations and red-teamed GPT-Live for risks specific to voice: self-harm, emotional reliance, violence, and sexual content.
Why it matters
150 million people use ChatGPT Voice weekly. The jump from turn-based to full-duplex is not a feature upgrade — it is a different modality. The difference between "query a search engine" and "talk to a colleague" changes what people use voice AI for.
For developers, the API pathway is opening. OpenAI has a signup form for developers who want GPT-Live access programmatically. When the API lands, the voice-agent space — currently fragmented across Deepgram, ElevenLabs, and bespoke pipelines — consolidates around a single provider running the most popular consumer voice product on the market.
Both models are filed in the directory as conditional. The full-duplex capability is real and shipping today — GPT-Live-1 earns a clear preference over Advanced Voice Mode on turn-taking, conversational flow, and naturalness in head-to-head human evaluations, plus strong gains on GPQA, BrowseComp, and τ³-Voice Telecom benchmarks (OpenAI, 2026(opens in new tab)). The conditions: no API access yet, and language support is uneven with accent and fluency gaps for some languages. Third-party latency benchmarks and API availability will determine when the verdicts move.
| Benchmark | GPT-Live-1 vs. AVM |
|---|---|
| GPQA (scientific reasoning) | Substantially outperforms Advanced Voice Mode |
| BrowseComp (agentic web search) | Strong gains over Advanced Voice Mode |
| τ³-Voice Telecom (voice agent tasks) | Outperforms Advanced Voice Mode |
What GPT-Live changes for you
If you are on ChatGPT Plus, Pro, or Go: GPT-Live-1 is your new voice model. It ships with configurable reasoning depth — Instant, Medium, or High effort — and supports visual response cards and real-time translation. The richer expressivity and higher reasoning ceiling are a material upgrade over Advanced Voice Mode.
If you are on the free tier: GPT-Live-1 mini delivers the same full-duplex architecture with GPT-5.5 Instant for background reasoning. It is a genuine step forward from the old turn-based voice mode, though it lacks the Medium/High reasoning effort tiers of the paid model.
If you are a developer: Sign up for API access. The programmatic interface is not yet available, but the signup form is live and OpenAI has confirmed it is coming. When it lands, expect the voice-agent market to consolidate rapidly around a provider with 150M weekly active voice users.
FAQ
Is this an AI companion? No. OpenAI's safety system red-teamed GPT-Live specifically for emotional reliance and parasocial risk — audio-native evaluations for self-harm, psychosis, and mania are part of the deployment. The model is engineered for conversation, not companionship.
How is this different from Advanced Voice Mode? Advanced Voice Mode was turn-based: you stopped speaking, then the model responded. GPT-Live is full-duplex: it listens and speaks simultaneously, handling interruption, turn-taking, and background reasoning concurrently. It is the architectural difference between a walkie-talkie and a phone call.
When will the API be available? OpenAI has not published a date. A developer signup form is live, and the product briefing confirmed programmatic access is coming. Until then, GPT-Live is ChatGPT-only.
Affected tools & models
Never need to catch up again
The weekly delta — only verdict changes and act-now items. No digest filler.