The Pulse
OpenAI Ships GPT-Live-1 Voice Model in Its API
OpenAI has released GPT-Live-1 in its API, giving developers a full-duplex voice model that can listen and speak at the same time. The model supports interruption handling, telephony, delegated reasoning and a front-end voice price of $0.05
AI.info Team ·
Interruptions fell by almost 80% in one early evaluation. OpenAI says that is what happened when language learners used GPT-Live-1 instead of a conventional turn-based voice system, where the assistant waits for a complete speaking turn before responding.
OpenAI released GPT-Live-1 in its API on September 10, 2026, bringing the voice model from ChatGPT to developers building customer-service, tutoring, telephony and other spoken applications. The model listens and speaks at the same time, while separate backend models can handle deeper reasoning, web searches and tool calls.
The release gives developers more control over how a voice assistant sounds and behaves, but it also marks a technical change in how OpenAI expects voice systems to be assembled. Rather than stitching together speech recognition, a language model and speech synthesis for every exchange, developers can use GPT-Live-1 as a continuous voice layer and connect it to the models and software systems behind it.
GPT-Live-1 keeps talking while other models work
Traditional voice agents process conversations as a chain: speech-to-text first, reasoning second and text-to-speech last. Each handoff can add delay or lose information about pauses, tone and interruptions. GPT-Live-1 instead reasons over incoming and outgoing audio together, allowing it to keep listening while it speaks and respond when a user changes direction.
OpenAI says the model can delegate harder work to a backend text model such as GPT-6 Astra or to a third-party model. That division lets GPT-Live-1 maintain the live conversation while another system searches, reasons through a complex request or calls a business tool. Developers can choose the backend model, tools and orchestration system rather than tying every part of the application to one model.
The API also provides automatic speech-recognition transcripts and response text, supports keyword biasing and includes native turn detection for applications that still need explicit conversation boundaries. OpenAI says GPT-Live-1 handles background noise and silence more quietly, avoiding unnecessary spoken commentary when a user pauses to think.
OpenAI reports a 30-point gain in full-duplex testing
OpenAI reports that GPT-Live-1 improved its Full Duplex Bench score by 30 percentage points over GPT-Realtime-2.1. The company says the largest gains came in turn-taking latency and interactive behavior, including pauses, interruptions, backchannel responses and speech occurring in the background.
In early testing by Speak, a language-learning company, GPT-Live-1 reduced interruptions during learners’ thinking pauses by almost 80% compared with previous turn-based systems. OpenAI also says GPT-Live-1 paired with GPT-6 Astra at medium reasoning effort ranked first on its Tau3 evaluation for end-to-end voice-agent tasks.
“Adding GPT-Live-1 into Yelp Host and Hatch improved turn-taking and accuracy over our traditional voice architecture. When Yelp Host uses GPT-Live-1 to answer calls, like reservations and food orders, we're seeing meaningful improvements in call handling rates.”
Alex Levy, chief technology officer, Yelp
The results come from OpenAI’s own evaluations and partner feedback, rather than an independent benchmark. They point to the specific problem the company is targeting: voice systems that can recognize when someone is still thinking, accept an interruption and continue a conversation without forcing the caller to wait through a complete response.
Telephony and a wider set of voices
GPT-Live-1 supports full-duplex voice agents for phone calls, including use cases such as restaurant reservations and customer support. OpenAI’s announcement names Yelp, Speak, Fin and Cognition among the companies using or evaluating the model.
The release also expands the available voice selection beyond a small group of real-time voices. Developers can choose from voices including Quartz, Ripple, Vesper, Willow, Stone, Gleam, Meridian, Bossa, Tempo, Beacon, Delta and Cinder, with OpenAI promising additional languages and voice options over the coming months.
OpenAI’s own examples include connecting GPT-Live-1 to Codex. In that setup, the voice model receives the user’s spoken request, sends relevant context to Codex and delivers the resulting answer back into the live conversation. The design separates the audio exchange from the longer-running software task, so a slow tool call does not have to stop the conversation.
The API starts at five cents per minute
GPT-Live-1 is available in the API at $0.05 per minute for the front-end voice layer. OpenAI says developers can pair it with the backend model and application stack that fit their product, meaning the published rate does not represent the full cost of a voice agent that also uses reasoning models, tools, telephony services or business systems.
OpenAI also offers a managed enterprise route through OpenAI Presence, which uses GPT-Live-1 for voice interactions that can answer questions, access company systems, take approved actions and escalate to human staff. Custom voice access requires contacting OpenAI’s sales team.
GPT-Live-1’s API release turns the model’s main ChatGPT feature into a developer platform: continuous conversation on the front end, deeper computation behind it and phone support when an application needs to reach users outside a chat window. The immediate test for OpenAI will be whether the lower-latency interaction and simpler voice layer translate into reliable calls at the price developers actually pay once the backend work is included.