GPT-Live-1: Full-Duplex Voice Model in the API
OpenAI brought GPT-Live-1 to the API on 10 September 2026 and says it brings natural, full-duplex conversations to voice apps: the model listens and speaks at the same time, smooths over interruptions and backchannels, and delegates deeper reasoning and tool calls to whatever backend model it is paired with (such as GPT-6 Astra or a third-party model); developers can shape tone, pace and style through the system prompt and natively receive transcripts and response text, keyword biasing and turn detection; OpenAI says it costs $0.05 per minute for the front-end voice layer and supports WebRTC, WebSockets and telephony.
Tool Overview
Features, steps and FAQ below
Features
- ✓ Full-duplex conversation: OpenAI says GPT-Live-1 listens and speaks at the same time, reasoning over incoming and outgoing audio in a single model and avoiding the latency and brittle handoffs of chained STT-LLM-TTS architectures
- ✓ Interruption handling: OpenAI says it responds naturally to pauses, interruptions and brief acknowledgements; in early evaluations the language-learning product Speak found it cut interruptions during thinking pauses by almost 80% versus previous turn-based systems
- ✓ Reasoning and tool-calling delegation: OpenAI says GPT-Live-1 can delegate reasoning and tool calls to a backend text model, for example using Luna for high-volume tasks and Astra for complex customer issues that require reasoning
- ✓ Tone, pace and style: OpenAI says developers can shape an agent's tone, pace and conversational style through the system prompt, with an expanded set of voices across accents, dialects and languages
- ✓ Long-session reliability and telephony: OpenAI says it improves context retention and conversational quality across extended interactions, better handles background noise and silence, and supports deploying full-duplex voice agents for phone calls
How to Use
- Read the GPT-Live-1 Live guide in the OpenAI developer docs and get API access
- Create a session and pick the backend model and agent framework that fit your product, letting GPT-Live-1 handle the conversation layer while the backend handles deeper reasoning and actions
- Set the agent's tone, pace and conversational style through the system prompt, and use keyword biasing and turn detection as needed
- Connect over WebRTC (browser apps), WebSockets (server integrations) or telephony; OpenAI says the front-end voice layer is billed at $0.05 per minute
FAQ
What is GPT-Live-1?
OpenAI says GPT-Live-1 is its natural, full-duplex voice model available in the API, able to listen and speak at the same time and to delegate deeper reasoning and actions to the models and tools it is paired with, for building voice-enabled apps and business workflows.
How is it billed?
OpenAI says the front-end voice layer costs $0.05 per minute, and it is paired with the backend model and agent framework you choose, whose token and tool costs are billed separately.
What are its key strengths?
OpenAI lists improved interruption handling through a single model that reasons over incoming and outgoing audio, reasoning and tool-calling delegation to a backend model, control over tone/pace/style via the system prompt, better handling of background noise and silence, long-session reliability, and telephony support.
How does it perform?
OpenAI says GPT-Live-1 improves Full Duplex Bench performance by 30 percentage points over GPT-Realtime-2.1 with large gains in turn-taking latency and interactive behavior, and that paired with GPT-6 Astra at medium reasoning effort it ranks #1 on Tau3, a benchmark for frontier voice-agent intelligence on end-to-end tasks.
Which companies are using it?
OpenAI says customers including Yelp, Speak, Fin and Cognition (Devin) are using GPT-Live-1, with Speak finding in early evaluations that it cut interruptions during thinking pauses by almost 80%.