OpenAI's GPT-Live-1 API: The Voice Model That Listens and Speaks at Once
Last updated: September 2026
Five cents per minute, and thirty percentage points over its predecessor on the leading duplex-conversation benchmark — those are the headline numbers behind GPT-Live-1, the voice model OpenAI opened to API developers on September 10, 2026. It is the company's first API model that listens and speaks in the same moment through a single-model architecture instead of a chained pipeline. For startups and enterprises building voice products, the arrival pushes natural, interruptible conversation from an expensive engineering project toward a metered utility.
Source: OpenAI — official announcement page
What is GPT-Live-1?
A fully full-duplex voice model: the listening channel and the speaking channel run concurrently, the way two humans actually talk. OpenAI first shipped this level of conversation to consumers inside ChatGPT — our coverage of that rollout is here: GPT-Live, OpenAI's Voice Mode in ChatGPT. What is new is that the same capability now sits in the API, meaning any developer or company can build it into their own products.
The architecture is the heart of the difference. Conventional voice systems are a relay of three stages: speech-to-text, a text model, then text-to-speech. GPT-Live-1 replaces the relay with one model reasoning over incoming and outgoing audio together, which is why it can be interrupted mid-sentence, hold its place, and resume naturally — and why it manages background noise and silence without narrating every step.
| Criterion | Traditional STT→LLM→TTS chain | GPT-Live-1 |
|---|---|---|
| Number of models | three chained components | one voice model |
| Listening while speaking | effectively impossible | native (full-duplex) |
| Natural interruption | noticeable delay | handled instantly by one model |
| The "brain" | the middle text model | delegated to a backend model like GPT-6 Astra |
| Noise and silence | chain degrades | natively managed |
The elegant split in the design is the separation of voice from reasoning. GPT-Live-1 owns the voice layer completely: conversational pacing, interruption handling, tone and style controlled through the system prompt, and reliability across long sessions. Logical reasoning and tool calls are delegated to a backend text model of your choice — GPT-6 Astra at medium reasoning effort is the pairing OpenAI highlights, and third-party models are supported too. We profiled the leading backend candidate in: OpenAI Launches GPT-6 Astra.
The numbers that matter to developers
| Benchmark | GPT-Live-1 result | Stated comparison |
|---|---|---|
| Full Duplex Bench | +30 percentage points | over GPT-Realtime-2.1 |
| Tau3 ranking | #1 | with GPT-6 Astra at medium reasoning effort |
| Interruptions (Early Speak) | down almost 80% | vs. previous turn-based systems |
| Customer-service evals | Pass@1 results | airline, retail, telecom |
| Banking knowledge | 97 tasks | banking-sector evaluations |
The announced pricing is $0.05 per minute for the front-end voice layer — the listening, speaking, and turn management — while custom brand voices require contacting sales. The backend model you delegate reasoning to is billed on its own standard pricing, and the reasoning-effort dial stays in your hands to balance quality against cost.
| Time unit | Front-end voice cost only |
|---|---|
| One minute | $0.05 |
| One hour | $3 |
| 1,000 minutes (~17 hours) | $50 |
| Contact center: 10,000 min/month | $500 |
That math is what turns "conversational AI" from a demo into a budget line a mid-size operation can defend in a meeting.
Under the hood: the developer surface
On the input side, the model ships native ASR transcripts, response text, and turn detection, plus alphanumeric understanding and keyword biasing — small features that decide whether a phone agent reliably captures an account number, a name spelling, or a sector-specific term. Voice options span accents, dialects, and languages, with more promised in the coming months.
Telephony support means agents that operate on ordinary phone lines, not only inside apps — a prerequisite for the largest volume use case on earth: call centers. And long-session reliability matters for exactly that workload, where a support conversation can stretch well past the demos.
Source: OpenAI — official announcement page
For the technical starting point, the official Live API guide on developers.openai.com is the entry document, and the model is also usable through OpenAI Presence.
Who should build on it first?
The three sectors OpenAI measured in its announcement — airline, retail, and telecom customer service — happen to be the three pillars of customer support volume worldwide. Add the banking-knowledge evaluation (97 tasks) and the picture is unmistakable: banks, telecoms, airlines, and delivery platforms are the natural first customers, running millions of voice minutes a month against strict quality bars.
Four build-ready use cases stand out:
A live phone support line. Telephony support means the agent answers a real phone call, understands the caller's phrasing, checks records through tool calls, and respects immediate interruptions when the caller changes their mind.
A hands-free internal assistant. Warehouse, lab, and vehicle workflows where employees update systems by voice while their hands stay busy, with the native transcripts providing an audit trail afterward.
A conversational language-learning platform. Full-duplex is transformative for learners: they can interrupt, be interrupted, and rephrase — while the tone controls shape a patient tutor's register, and alphanumeric capture cleans up spelling drills.
An appointment intake front desk. The underrated feature is silence handling: the model waits and asks rather than collapsing when a caller pauses to find their reference number — ordinary human behavior that breaks lesser systems.
In all four, the same engineering rule applies: launch narrow (ten question types, say), measure the resolution rate without human intervention, then expand. The swappable backend means you can upgrade the "brain" without rebuilding the "voice."
Frequently asked questions
What is OpenAI's GPT-Live-1 model?
A full-duplex voice model launched in the API on September 10, 2026. It listens and speaks simultaneously through a single model rather than a chain of conversions, delegating deep reasoning to a backend text model the developer chooses, such as GPT-6 Astra.
How much does GPT-Live-1 cost per minute?
Five US cents per minute for the front-end voice layer — three dollars per hour — while backend model invocations are billed at their own standard rates. Custom brand voices require a sales agreement.
How is GPT-Live-1 different from GPT-Realtime?
GPT-Live-1 scores thirty percentage points above GPT-Realtime-2.1 on the Full Duplex Bench and cuts failed interruptions by roughly 80% versus previous turn-based systems, thanks to its single-model architecture managing incoming and outgoing audio together.
Does GPT-Live-1 support phone calls?
Yes — telephony support is built in, so agents can operate on standard phone lines rather than only inside applications, which makes high-volume call-center deployments practical.
Which backend model does GPT-Live-1 use?
You choose it. The announcement highlights GPT-6 Astra at medium reasoning effort as the reference pairing, and third-party models are also supported, with reasoning effort as a cost-quality dial.
Sources
- OpenAI — Introducing GPT-Live-1 in the API — the complete official announcement text
- Live API developer guide — technical documentation for getting started
- OpenAI — Terms of Use — the governing terms for commercial use
What to watch in the coming weeks
Three developments are likely: a longer published list of supported languages and dialects, clearer volume pricing as adoption matures, and direct competition from other duplex-conversation engines. What is already certain is that dropping natural voice to five cents a minute moves the "voice employee" from lab pilots into real budgets — and the earliest durable wins usually go to teams that experimented before the market crowded. If you build for students and researchers and want a platform that already aggregates the ecosystem — scholarships, rankings, study tools — Truescho is free to try.