OpenAI's GPT-Live splits the voice model from the reasoning model

What GPT-Live actually changes about the voice stack

OpenAI describes the original ChatGPT Voice as a cascade of three models: speech-to-text to transcribe you, a large language model to produce a response, and text-to-speech to speak it back. Each turn passed sequentially through the chain.

GPT-Live replaces that pipeline with what OpenAI calls a full-duplex architecture — one that listens and speaks at the same time rather than waiting for a completed turn. The company names two specific failure modes of the cascade it is trying to fix: information lost as text moved between models, and responses that were slow and stilted.

The practical tell is backchanneling. OpenAI says GPT-Live can interject 'mhmm' or 'yeah', hold a quick back-and-forth, or stay quiet while you think. Those are timing behaviors, not smarter answers — they only work if the model is processing audio continuously instead of turn by turn.

The delegation move: a fast voice model calling a slow frontier model

The most consequential design choice is that GPT-Live does not try to be the frontier model. For questions requiring web search, deeper reasoning, or more complex work, it delegates to a separate model behind the scenes and brings the result back into the conversation when it's ready.

At launch that background model is GPT-5.5, and OpenAI says it will swap in newer frontier models over time without changing the voice layer. Crucially, GPT-Live keeps talking to you while the delegated work runs — the conversation flow is decoupled from the latency of the reasoning call.

This is a clean separation of concerns: one model owns the real-time audio interaction, another owns the hard thinking. It's a different bet than a single monolithic multimodal model handling both, and it lets OpenAI upgrade intelligence and interaction independently.

Two sizes, ChatGPT first, API later

OpenAI is shipping two variants, GPT-Live-1 and GPT-Live-1 mini, rolling out to ChatGPT users globally starting the day of the announcement. The mini variant signals that OpenAI expects developers to trade off cost or latency against capability once the models are available more broadly.

API access is described as coming 'soon,' with a sign-up form for developers and enterprises. So the immediate release is a consumer ChatGPT Voice upgrade; the platform play for builders is deferred.

OpenAI frames the longer-term ambition explicitly: it believes this research will 'unlock the ability to use voice for increasingly complex, longer-running, and more agentic work.' That points past chat toward voice as an interface for tasks that run in the background.

What the delegation pattern implies for voice frontends

For teams building voice interfaces, the notable idea here is that conversational responsiveness and answer quality are now handled by different components with different latency budgets. A frontend can feel present and human while the substantive work happens asynchronously.

That reframes what a voice UI has to do. Instead of forcing users to wait through a single slow turn, the interaction layer can acknowledge, hold context, and narrate progress — then surface the frontier model's result when it lands. The hard problem shifts from 'answer fast' to 'stay engaged while the real answer is being computed.'