Skip to content
NewsModel release

OpenAI ships GPT-Live-1, a full-duplex voice model that can listen and speak at the same time

· by Pondero Newsdesk

The short version

OpenAI released GPT-Live-1 and GPT-Live-1 mini on July 8, replacing Advanced Voice Mode with models that process incoming and outgoing audio simultaneously, while routing complex queries to GPT-5.5 in the background.

OpenAI ships GPT-Live-1, a full-duplex voice model that can listen and speak at the same time

Every voice AI product since 2023 has shared one constraint: the model stops listening when it starts speaking. OpenAI dropped that assumption on July 8, 2026, releasing GPT-Live-1 and GPT-Live-1 mini under a full-duplex architecture that processes incoming and outgoing audio in parallel.

What

The previous ChatGPT Voice pipeline chained three separate models: speech-to-text, then a large language model, then text-to-speech. Advanced Voice Mode (2024) improved on that by collapsing the chain into a single model, but it remained turn-based, meaning it waited for a pause in the user's audio before responding. Per OpenAI's announcement, GPT-Live-1 makes interaction decisions many times per second rather than once per turn: whether to speak, continue listening, pause, interrupt, or invoke a tool.

For queries that require search, deep reasoning, or agentic execution, GPT-Live-1 delegates to GPT-5.5 in the background and keeps the conversation moving while that work completes. OpenAI describes four variants: GPT-Live-1 Instant and GPT-Live-1 mini route to GPT-5.5 Instant; GPT-Live-1 Medium and High route to GPT-5.5 Thinking at increasing reasoning effort.

Rollout began July 8 globally on iOS, Android, and ChatGPT.com. GPT-Live-1 is the new default for Go, Plus, and Pro subscribers; free users get GPT-Live-1 mini. Per OpenAI, the models come to the API "soon," with a developer sign-up form live now. Video and screen-sharing are not supported at launch, and the legacy Advanced Voice Mode stays accessible during the transition.

OpenAI said in a press briefing that more than 150 million people use ChatGPT Voice or Dictation each week, per TechCrunch.

Why it matters

The routing split is the structural change worth understanding. Keeping a lightweight real-time model in charge of the voice channel while delegating heavy reasoning to GPT-5.5 only when needed lets OpenAI deliver sub-second conversational responses without sacrificing accuracy on hard questions. Developers building voice agents have had to wire that tradeoff by hand, with custom intent-detection and parallel request plumbing. GPT-Live-1 handles it natively.

Two constraints matter for builders right now. The API is not live at launch, so no production integrations can run yet. Voice-plus-video, which underpins hands-free agent prototypes, also arrives later.

OpenAI published a system card at deploymentsafety.openai.com/gpt-live covering safety evaluations, real-time output safeguards, and age-appropriate behavior for teen users. That is worth reading before any enterprise evaluation: always-on audio processing raises different disclosure and data-handling questions than a text interface.

Per OpenAI, GPT-Live-1 substantially outperforms Advanced Voice Mode on GPQA (expert-level scientific reasoning across biology, chemistry, and physics) and shows strong gains on BrowseComp (agentic web search). Human evaluations across 5-to-10-minute conversations on turn-taking, interruptions, and conversational flow showed GPT-Live-1 was "strongly preferred" over Advanced Voice Mode. These are vendor-run evaluations, not third-party audits.

What to watch next

The API launch date and pricing are the near-term unknowns. Third-party latency benchmarks against Google's real-time voice stack will clarify where GPT-Live-1 fits in production cost-performance trade-offs. The live-translation capability that OpenAI demoed also showed a pronounced American accent in Hindi during the briefing; broader language quality data will matter for any non-English deployment.

Sources