OpenAI has released GPT-Live-1 in its API, bringing the model’s natural, full-duplex voice behavior to developers building assistants, phone workflows, and other spoken interfaces. The company describes the release as a voice layer that can listen and speak at the same time, rather than waiting for one side of a turn to finish before processing the other.
A shorter path from speech to action
Traditional voice agents are usually assembled as a chain: speech recognition turns audio into text, a language model reasons over that text, and text-to-speech generates a reply. OpenAI says GPT-Live-1 can reason over incoming and outgoing audio together, which should reduce handoffs that add latency or lose the rhythm of a conversation. It also supports natural interruptions, pauses, acknowledgements, and background noise.
That does not mean the model must do every job itself. OpenAI says GPT-Live-1 can delegate deeper reasoning and tool calls to a backend text model, including GPT-6 Astra or a third-party model. In practice, a voice agent could use a faster backend for routine scheduling and a more capable one for complex customer issues while keeping the spoken interaction continuous.
What OpenAI is measuring
In OpenAI’s reported evaluations, GPT-Live-1 improved Full Duplex Bench performance by 30 percentage points over GPT-Realtime-2.1. The company also says that GPT-Live-1 paired with GPT-6 Astra at medium reasoning effort ranked first on Tau3, an end-to-end benchmark for voice-agent intelligence. These are vendor-reported results, so they are best read as evidence of the product’s intended operating profile rather than as an independent industry ranking.
OpenAI also cites an early evaluation from language-learning company Speak, where the company says GPT-Live-1 cut interruptions during learners’ thinking pauses by almost 80% compared with previous turn-based systems. Customer examples from Yelp, Fin, and Cognition point to the same practical goal: agents that feel less like a sequence of API handoffs and more like a live conversation.
The interface is configurable, not one-size-fits-all
Developers can steer tone, pace, and conversational style through the system prompt. The model natively provides speech-recognition transcripts and response text, supports keyword biasing and alphanumeric understanding, and retains turn detection for applications that still need explicit turn boundaries. OpenAI is also expanding the set of voices across accents, dialects, and languages.
Availability and the practical trade-off
GPT-Live-1 is available in the API, with OpenAI listing a price of $0.05 per minute for the front-end voice layer. Deeper reasoning, tools, telephony, escalation, and enterprise context still sit around that layer and carry their own product and operational choices. The announcement makes the voice loop simpler; it does not remove the need for consent, logging, human handoff, or careful evaluation in high-stakes calls.
The broader signal
The important shift is architectural. OpenAI is treating spoken interaction as a native model capability while leaving the application’s reasoning stack modular. If the approach holds up outside the company’s benchmarks, voice developers may spend less time repairing turn-taking and more time designing the work an agent is allowed to do.