Voice software has spent years pretending conversation is a queue. One person talks. The system waits. The system talks. The person waits.
While that pattern works on a help-line menu, it breaks the moment someone talks over a sentence, adds a detail mid-thought, or goes quiet because they are thinking. Call centers, language tutors, and booking lines all run into the same gap: the model that sounds fluent still treats speech like a form.
OpenAI wants to solve this particular issue by releasing 'GPT-Live-1, a voice model that already powers ChatGPT Voice for paid users, into the developer platform.
The company first shipped the Live family inside ChatGPT in July. The September 10 launch is the version built for other people's products: phone agents, support lines, tutors, and apps that need a spoken front end while a separate model does the work.
Sessions are priced at $0.05 per minute for the voice layer, billed by the second. Reasoning, tools, and backend models are charged on their own meters.
The design is a split.
GPT-Live-1 handles listening and speaking in one model. It can keep talking while it hears the next sentence, pick speech out of café noise, and change direction before it finishes a turn.
OpenAI says the model can mirror a speaker's emotion and speed.
For example, it understands humans' dynamic elements of spontaneous speech.
Users can interrupt, or speak over the system, or even change the topic mid-sentence while the system stops talking instantly and adapts. Users can self-correct, or modify their query without having to break the interaction. It even understands natural human speech behaviors like laughter, hesitations, short acknowledgments like "yeah" or "gotcha," and briefly turning to speak to someone else before resuming.
Deeper work is delegated.
Developers can send that work to GPT-6 Astra, Codex, another OpenAI model, or a third-party harness.
The conversation does not have to freeze while a reservation is checked or a ticket is looked up.
OpenAI's older Realtime stack often asked developers to chain speech recognition, a language model, and speech synthesis. Live-1 is meant to collapse that voice layer so the brittle handoffs sit elsewhere, or disappear.
Developers can set tone, pace, expressiveness, language, and response length in instructions.
The API release also adds twelve realtime voices with names such as Quartz, Ripple, Vesper, Willow, Stone, Gleam, Meridian, Bossa, Tempo, Beacon, Delta, and Cinder.
Telephony is part of the pitch, from restaurant reservations to support queues.
The model returns ASR transcripts and response text, supports keyword biasing, and still exposes turn detection for teams that want explicit boundaries even though the model itself is not turn-based.
Image and video are not supported.
Rate limits are counted in concurrent sessions, from 25 on Tier 1 to 500 on Tier 5. The free tier is not supported. Custom voices go through sales. OpenAI Presence is another path that uses Live-1 for realtime voice.
Audio from GPT-Live has carried SynthID watermarking since a July update, with a public verification tool and API checks for provenance.
The numbers OpenAI published are comparisons against its own prior voice model, GPT-Realtime-2.1, not a public bake-off against every competitor.
Full Duplex Bench rose by 30% points, with the largest gains in turn-taking latency and interactive behavior.
Community notes from the developer announcement put turn-taking latency at 0.798 seconds against 1.41 seconds for Realtime-2.1.
Paired with GPT-6 Astra at medium reasoning effort, Live-1 completed 83.6% of Tau3 tasks on the first attempt, against 45.7% for Realtime-2.1.
Separate Full Duplex Bench v3 figures cited there include 87% tool-calling success and 90% response quality.
[x]https://x.com/juberti/status/2098263283182817465[x]
Those are vendor evaluations. They describe a system tuned for overlap, pause, and course correction rather than a single longer monologue.
The competitive field is the phone stack that companies already run: older realtime models, cascaded speech pipelines, and in-house IVR.
Live-1 does not remove the need for a backend that knows the business. It changes where that backend sits. The voice model keeps the floor.
The reasoning model books the table, files the ticket, or writes the code.
