OpenAI announced Thursday that it is bringing GPT-Live-1, the full-duplex voice model behind ChatGPT's natural voice experience, to its API — letting developers build applications that listen and speak at the same time, without stitching together speech-to-text, a reasoning model, and text-to-speech by hand.
The model first launched in July as the engine for ChatGPT Voice. Now that it is open to third parties, OpenAI has added developer-facing capabilities: shaping an agent's tone, pace, and conversational style through the system prompt, and better handling of background noise and silence without interrupting the user or narrating every step out loud. Developers choose the backend reasoning model themselves — Luna for high-volume tasks like scheduling, Astra for complex customer issues — matching depth, speed, and cost to each job.
Traditional voice agents are chained: record audio, transcribe it, reason, synthesize speech, play it back. Every handoff adds latency and makes "interruption" a hard engineering problem. GPT-Live-1 handles listening and speaking in a single model, catching interruptions as they happen, waiting through pauses, and delegating deeper reasoning and tool calls to a backend text model while the conversation continues. In early evaluations, language-learning app Speak found GPT-Live-1 cut interruptions during thinking pauses by almost 80% versus previous turn-based systems.
In OpenAI's evaluations, GPT-Live-1 improves Full Duplex Bench performance by 30 percentage points over GPT-Realtime-2.1, and paired with GPT-6 Astra at medium reasoning effort it ranks #1 on Tau3, a frontier voice-agent intelligence benchmark. Pricing is $0.05 per minute, billed per second, with backend model and tool usage billed separately. The model natively provides ASR transcripts and response text, supports strong alphanumeric understanding and keyword biasing, and enables telephony use cases from restaurant reservations to customer support.
Customers are already reporting results. Yelp uses it to answer reservation and food-order calls, seeing meaningful improvements in call handling rates and fuller, more natural sentences from callers. Speak says it "knows when to give learners space and when to step in." Fin, an enterprise support platform, says it moves voice support "from the stop-start rhythm toward the natural flow of a phone call." OpenAI also shipped a Codex integration example and OpenAI Presence, an enterprise product that uses GPT-Live-1 for real-time voice agents that query company systems, take approved actions, and escalate to people when needed.
From product to platform, the barrier to voice AI is falling. When interruptions, overlapping speech, and noisy environments are digested naturally by the model, voice may become, faster than we expect, the most ordinary way we talk to machines.