OpenAI's GPT-Live-1 voice API is a full-duplex voice model that listens and speaks at the same time — it handles pauses, interruptions and backchannels mid-sentence, runs at 0.05 US dollars per minute billed by the second, and, per OpenAI's own launch figures, nearly doubles task completion over the previous GPT-Realtime-2.1 voice model. The announcement went out on 10 September 2026 on OpenAI's official developer forum under the title "Introducing GPT-Live-1 in the API", and it is the company's first API model built around simultaneous listening and speaking rather than the old take-turns pattern. If you build voice agents — phone assistants, booking bots, support lines — this release changes what a natural conversation can feel like.
📺 Watch: GPT 6 Astra + Hermes Voice Agent is a GAMER CHANGER! 🤯
🔥 Get the Agent OS as a free bonus: AI Profit Boardroom members get the full Agent OS zip, prompt libraries, daily tutorials and weekly live coaching calls. → Get inside · Want AI SEO help 1-on-1? Book a free SEO strategy session →
What the GPT-Live-1 Voice API Actually Does
The defining feature is full duplex. Older voice pipelines worked like a walkie-talkie: the system recorded you, transcribed the audio, ran it through a language model, synthesised a reply, and only then started talking — and while it talked, it was deaf. GPT-Live-1 processes input and output audio together, so the model can keep listening while it speaks. According to the 10 September 2026 launch post, it "listens and speaks simultaneously, handles pauses, interruptions, and backchannels" — those little "mm-hm" and "right" noises humans make constantly that used to derail voice bots entirely.
Alongside the duplex core, the launch post lists native ASR transcripts, explicit turn detection, and improved handling of silence and background noise. Turn detection is the underrated one: knowing when a caller has actually finished talking, versus merely pausing to think, is most of what separates a natural phone agent from an infuriating one.
GPT-Live-1 Benchmarks vs GPT-Realtime-2.1
OpenAI published direct comparisons against its previous voice model in the launch post, and the gaps are large. All figures below are OpenAI's own reported numbers from the 10 September 2026 announcement:
| Metric | GPT-Live-1 | GPT-Realtime-2.1 |
|---|---|---|
| Task completion (Tau3) | 83.6 percent | 45.7 percent |
| Turn-taking latency | 0.798 seconds | 1.41 seconds |
| Tool-calling success | 87 percent | not stated |
| Full-duplex interactivity | 80.1 percent | not stated |
The Tau3 jump — from 45.7 to 83.6 percent task completion — is the number that matters for anyone shipping agents that actually do things, because it measures whether the voice model finishes multi-step jobs rather than just chatting pleasantly. Sub-second turn-taking latency is the other headline: at 0.798 seconds, the response gap sits inside the range of ordinary human conversation. Treat vendor benchmarks as vendor benchmarks, of course — but these are at least concrete, named metrics rather than vibes.
If you want to build AI agents that answer calls, book appointments and handle customers around the clock, check out the AI Profit Boardroom — get the agent workflows and weekly coaching here. Want a personal plan for turning AI into traffic and revenue first? Book a free SEO strategy session with Julian and get it mapped out 1-on-1.
📺 Watch: Elevenlabs MCP + GPT-6 Astra + Hermes is WILD!
Voices, Style Control and How You Connect
GPT-Live-1 ships with 12 voices, named in the announcement as Quartz, Ripple, Vesper, Willow, Stone, Gleam, Meridian, Bossa, Tempo, Beacon, Delta and Cinder, spanning different accents and delivery styles. Beyond picking a voice, the launch post says you can steer tone, pace, speaking style and conversational behaviour through system prompts — so the same voice can play a brisk receptionist for one product and a patient support agent for another without any custom audio work.
Connection options cover the three places a voice agent actually lives. Per the announcement, developers connect over WebRTC for in-browser experiences, WebSockets for server-side audio streams, and Telephony or SIP for phone agents. That SIP path is the significant one for businesses: it is the route by which a voice model ends up answering a real phone number, which is where projects like the Telegram AI agent build and the Hermes voice agent setup have been heading — assistants that live where your customers already are rather than inside a demo page.
What the GPT-Live-1 Voice API Costs
Pricing is 0.05 US dollars per minute of conversation, billed per second, per the 10 September 2026 launch post — with the important caveat that backend model and tool usage are charged separately. GPT-Live-1 handles the talking; whatever agent brain you run behind it to look up orders, query calendars or write records bills on its own meter. A rough anchor: an hour of continuous conversation is 3 US dollars on the voice side before backend costs. Per-second billing also means short calls cost genuinely little, which suits the bursty traffic pattern of most small-business phone lines.
Where GPT-Live-1 Fits in the Current AI Stack
GPT-Live-1 arrives in a September 2026 lineup where OpenAI's text-and-agent flagship is GPT-6 Astra — covered in the GPT-6 Astra guide — and where the practical pattern for serious builds is a voice layer in front and an agent brain behind. The launch post's note that the model can keep a conversation going while a backend agent works is exactly that architecture: the voice model buys time gracefully while the real work happens elsewhere. That backend brain does not have to be OpenAI's; the Goldie Bench write-up covers how the main agent models compare in hands-on tests, and the GPT-5.5 Pro breakdown covers the previous generation if you are weighing older, cheaper options. Whichever brain you pick, giving it durable goals, memory and skills is the job of an operating layer like the Agent OS — a voice front-end is only as useful as the system it is wired into.
What to Build With It First
The obvious first builds for the GPT-Live-1 voice API are the ones where turn-taking quality is the whole product:
- Reception and booking lines — answer, qualify, book, hand off to a human when it matters, with the SIP connection putting it on a real number.
- Support triage — the interruption handling means callers can cut in with "no, not that order" without breaking the flow.
- Internal assistants — a spoken interface over the same agents you already run for research and admin, in the style of the personal assistant agent setups.
- Voice-driven content workflows — pair a talking front-end with production pipelines like the Hermes AI video generator workflow for hands-free operation.
For tool-heavy builds, the 87 percent tool-calling success figure from the launch post is the one to watch in your own testing: it is high for a voice model, but your agent's reliability will live and die on that number, so measure it on your actual tools — connected via something like the MCP server pattern — before you put it in front of customers.
How Full Duplex Changes the Way You Design an Agent
Designing for the GPT-Live-1 voice API is different from designing for a take-turns model, and it is worth internalising why before you build. With a half-duplex pipeline, your conversation design had to defend against dead air: you wrote short agent turns, avoided long lookups mid-call, and accepted that a caller interrupting would usually be ignored until the agent finished its sentence. Full duplex removes those constraints one by one. Because the model listens while speaking, a caller can cut in with a correction and the agent can adjust rather than ploughing on. Because the launch post describes explicit turn detection and better handling of silence and background noise, a thinking pause from the caller no longer triggers a premature answer. And because the model can keep the conversation alive while a backend agent works, a slow database lookup becomes a natural "let me just check that for you" moment instead of ten seconds of suspicious silence.
The practical redesign checklist that falls out of this: let agent turns breathe rather than clipping everything short, move heavy tool work behind conversational cover, and test your flows specifically against interruptions and backchannels — the situations the older generation of voice agents failed. The benchmark gap OpenAI reports between GPT-Live-1 and GPT-Realtime-2.1 suggests these are precisely the behaviours the model was trained to survive.
Is the GPT-Live-1 voice API available now?
Yes. Per the 10 September 2026 announcement on OpenAI's developer forum, GPT-Live-1 is live in the API with WebRTC, WebSocket and Telephony or SIP access. It is an API product, so you will need an OpenAI platform account with billing set up.
How is it different from GPT-Realtime-2.1?
GPT-Realtime-2.1 was OpenAI's previous real-time voice model. GPT-Live-1 replaces the take-turns interaction pattern with full duplex, and OpenAI's launch figures show 83.6 percent versus 45.7 percent task completion on Tau3 and roughly half the turn-taking latency at 0.798 seconds versus 1.41.
If you want working voice and agent systems — with the templates, the daily tutorials and 3,000-plus builders sharing what converts — check out the AI Profit Boardroom and get inside here. And if you want expert eyes on your own AI growth plan, book a free SEO strategy session — it costs nothing and you leave with a roadmap.
