OpenAI has launched GPT‑Live‑1 in its API, offering developers a real‑time, full‑duplex voice model for building voice‑enabled applications and business workflows. First introduced in ChatGPT, GPT‑Live‑1 can listen and speak simultaneously and can delegate deeper reasoning and tool use to backend models and systems.
Why this matters
Traditional voice agents chain together speech‑to‑text, a reasoning model, and text‑to‑speech, and each handoff adds latency and risks losing conversational timing or context. GPT‑Live‑1 handles listening and speaking within a single model, allowing it to react to interruptions and acknowledgements as they occur while offloading more complex tasks to backend models so the conversation can continue while work happens in the background.
Developers decide which models, tools, and agent harness to pair with GPT‑Live‑1. For example, a Luna‑like model can handle high‑volume routine tasks such as scheduling or order updates, while an Astra‑like model can be used for complex customer issues requiring deeper reasoning. This lets teams tune speed, cost, and reasoning depth to each task.
Early results and performance
In early evaluations, Speak found that GPT‑Live‑1 gave learners more time to think before the language tutor responded, reducing interruptions by almost 80% compared with prior turn‑based systems.
On benchmark tests, GPT‑Live‑1 improved Full Duplex Bench performance by 30 percentage points over GPT‑Realtime‑2.1, with significant gains in turn‑taking latency and interactive behavior. When paired with GPT‑6 Astra at medium reasoning effort, the combination ranked #1 on Tau3, a measure of frontier voice‑agent intelligence on end‑to‑end tasks.
The evaluations covered spoken customer‑service tasks in airline, retail, and telecom domains (using Pass@1 to measure task success with equal weighting across domains), spoken banking support with knowledge retrieval and account tools (Pass@1 over 97 banking_knowledge tasks), pause handling and turn taking, reactions to background speech and listener backchannels, reply latency (how quickly the agent starts a reply after the user finishes), and tool use from spoken requests that include natural pauses, hesitations, and self‑corrections.
Features and developer controls
- Native ASR transcripts and response text provided.
- Strong alphanumeric understanding and support for keyword biasing.
- Although not strictly turn‑based, GPT‑Live‑1 natively supports turn detection so developers can still design around explicit turn boundaries.
- Handles interruptions, backchannels, and other conversational dynamics.
- Tested for tool calls and spoken answers in scenarios with pauses, hesitations, and self‑corrections; Pass@1 scores measure tool‑call sequence and final answer alignment with reference intent.
Voices, languages and availability
OpenAI is expanding beyond a small set of real‑time voices to a broader selection across accents, dialects, and languages; voice and language availability will continue to grow in the coming months. For custom voice access, developers are directed to contact OpenAI sales to learn about eligibility and the request process.
Pricing and how to get started
GPT‑Live‑1 is available in the API today. The front‑end voice layer is priced at $0.05 per minute. Developers can pair GPT‑Live‑1 with the backend model(s) and agent harness that best fit their product and build voice experiences that scale with the work required.
Integrations and related offerings
The announcement notes example integrations such as connecting GPT‑Live‑1 to Codex, where an application passes conversation context to Codex and returns its answer to GPT‑Live‑1; setup and delegation handling are left to the developer. OpenAI also highlights OpenAI Presence as another way to build voice workflows on top of GPT‑Live‑1—Presence is positioned for enterprises that need trusted real‑time agents able to answer questions, resolve issues, use company systems, perform approved actions, and escalate to humans when necessary.
The public demo is time‑limited; using it implies agreement with OpenAI’s Terms and acknowledgement of the Privacy Policy.
