Overview
GPT-Live-1 is OpenAI's full-duplex speech model introduced to the API on September 10, 2026, after debuting in ChatGPT. A single model handles both speech input and output, listening and speaking at the same time and responding to interruptions and backchannels in real time, which is the fundamental difference from the traditional cascade of speech-to-text plus LLM plus text-to-speech.
In cascaded pipelines every handoff adds latency and loses timing, context or conversational rhythm, while GPT-Live-1 reasons over listening and speaking within one model and natively supports turn detection, so developers can still build around explicit turn boundaries. It can also delegate deep reasoning and tool calls to backend text models such as GPT-6 Astra, keeping the voice layer light while heavy work happens behind it.
According to the announcement, it improves Full Duplex Bench by 30 percentage points over GPT-Realtime-2.1 and ranks first on Tau3 with GPT-6 Astra as the backend. Language-learning app Speak measured nearly 80% fewer interruptions during thinking pauses, and one customer simplified its codebase by 80% and deleted 23,000 lines of code after migrating.
Key Features
- Full-duplex real-time interaction: A single model reasons over incoming and outgoing audio simultaneously, handling interruptions, backchannels and pauses naturally without cascaded handoffs.
- Delegated reasoning and tools: The voice layer can delegate deep reasoning and tool calls to backends like GPT-6 Astra, or pair with models like Luna for high-volume tasks such as scheduling and order updates, while conversation continues in the foreground.
- Controllable tone and style: Developers customize the agent's tone, pace and conversation style via system prompt, with 12 new voices covering more accents, dialects and languages.
- Silence and noise management: Better handling of background noise and silence without narrating every background step, with markedly improved context retention and conversation quality over long sessions.
- Native transcription output: Emits ASR transcripts alongside response text, with strong alphanumeric recognition and keyword biasing support for compliance and retrieval.
- Telephony support: Supports full-duplex phone voice agents, validated in real call scenarios such as restaurant reservations and customer service by products like Yelp Host.
Use Cases
- Voice customer service and reservations: full-duplex phone agents answer booking and ordering calls while callers speak in fuller, more natural sentences
- Language-learning partners: Speak measured nearly 80% fewer interruptions during thinking pauses, making conversations feel human
- Real-time voice assistants: any scenario that requires listening while speaking and recovering smoothly from interruptions
- High-volume task routing: scheduling and order updates handled by the voice layer with a lightweight backend, escalating hard questions to reasoning models
- Enterprise voice front-ends: OpenAI Presence drives real-time voice interaction on GPT-Live-1 for enterprise deployments
Pros
- The full-duplex architecture removes the multi-hop latency and handoff loss of cascaded pipelines
- 30 percentage points better on Full Duplex Bench than GPT-Realtime-2.1 and first on Tau3 with GPT-6 Astra
- Reasoning and tool calls delegate to backend models, decoupling the voice layer from the intelligence layer
- Customer migrations cut codebases by 80% and deleted 23,000 lines, greatly lowering maintenance cost
- 12 new voices cover more accents and dialects, with prompt-controllable tone and style
- Native transcript output makes compliance auditing and keyword retrieval straightforward
Pricing
The voice front-end layer costs $0.05 per minute; backend models and agent frameworks are chosen by the developer and billed separately. Custom voices and OpenAI Presence enterprise deployments require contacting sales.
Summary
GPT-Live-1 moves voice interaction from turn-based to full-duplex, making the most natural parts of human conversation, interruptions and backchannels, actually work in a product for the first time. Combined with delegable reasoning backends, it defines a new speech-agent architecture: the front end handles listening and speaking while the back end thinks. Teams building voice support, tutoring or real-time assistants should evaluate it seriously.
Version History
- GPT-Live-1 arrives on the API, opening the full-duplex voice model to developers (2026-09-10): The full-duplex voice model that debuted in ChatGPT is now on the API: a single model listens and speaks simultaneously, responds to interruptions in real time, and delegates deep reasoning and tool calls to backends like GPT-6 Astra. Full Duplex Bench improves 30 points over GPT-Realtime-2.1, the voice layer costs $0.05 per minute, and 12 new voices ship alongside.