OpenAI Ships GPT-Live-1: Full-Duplex Voice in One Model at $0.05 a Minute

Primary source: OpenAI — Introducing GPT-Live-1 in the API (September 10, 2026).

OpenAI put ChatGPT’s full-duplex voice model in the developer API

OpenAI launched GPT-Live-1 in the API on September 10, 2026 — the same full-duplex voice stack used in ChatGPT, now available for developers who need a model that can listen and speak at the same time.

Unlike turn-based voice pipelines, GPT-Live-1 is built for interruption handling and natural conversation flow. Developers can steer tone, pace, and speaking style through the system prompt, and optionally delegate heavier reasoning or tool use to backend models while the voice layer stays live.

Billing for the voice layer is about $0.05 per minute, metered per second. Backend model and tool usage is billed separately. Sessions run over the Live sessions API (WebSocket / WebRTC).

For operators, this is a voice-layer product story — a priced duplex endpoint you can wire into tutoring, support, or coaching apps — not a claim that every phone queue is solved.

What OpenAI published: Speak, Full Duplex Bench, Tau3

Named outcomes from the OpenAI post:

• Speak (language tutor): reported roughly 80% fewer interruptions versus a prior turn-based setup — fewer false cut-ins when a learner is still speaking.
• Full Duplex Bench: OpenAI reports about a 30 percentage point gain versus GPT-Realtime-2.1.
• Tau3: GPT-Live-1 ranked #1 when paired with GPT-6 Astra at medium reasoning.

Product shape in brief:

• One listen-and-speak model in the API (full duplex).
• Interruption handling as a first-class behavior, not a bolt-on VAD hack.
• Style control via system prompt; optional handoff of reasoning/tools to backend models.
• Live sessions API over WebSocket/WebRTC; voice billed ~$0.05/min per second.

Treat those numbers as vendor-published benchmarks and one customer-style tutor result, not as a guarantee for your IVR or call-center stack.

What this proves — and what it does not

Proves (so far):

• OpenAI is shipping ChatGPT’s full-duplex voice capability as a developer API product (GPT-Live-1), with a clear ~$0.05/min voice-layer price and separate backend billing.
• They are willing to publish concrete deltas: Speak ~80% fewer interruptions vs prior turn-based; ~30pp Full Duplex Bench gain vs GPT-Realtime-2.1; Tau3 #1 with GPT-6 Astra at medium reasoning.
• The architecture allows a thin voice layer plus optional delegation to heavier backend models for tools/reasoning — a split many ops teams already want.

Does not prove:

• That every tutoring, sales, or support voice bot will cut interruptions ~80% — Speak is one reported setup, not a multi-industry RCT.
• End-to-end cost of a production agent (voice + backend model + tools + telephony) from the $0.05/min line alone.
• Latency, compliance, recording, or PII controls for regulated call centers — those are outside the launch post’s scope.
• That Full Duplex Bench or Tau3 scores equal customer satisfaction on your tasks without your own baselines.

Treat this as a named product launch with published bench and tutor metrics — useful architecture pressure, not a universal voice-ROI guarantee.

Operator filter: what to do with a $0.05/min duplex headline

Use a short filter before you rewrite your voice roadmap around GPT-Live-1:

1. Is duplex actually the bottleneck? If your pain is wrong answers or missing CRM context, a better listen/speak model will not fix grounding.
2. What is the before/after map? Name the task (tutor turn, support intake, coaching loop) and measure interruptions, talk-over, and completion — not “voice feels natural.”
3. Where does reasoning live? Voice layer vs backend model vs tools. GPT-Live-1’s optional delegation only helps if you already know which steps stay on the wire.
4. What is the all-in minute? $0.05/min voice + backend tokens + telephony + human escalation. Price the stack, not the slogan.
5. Who owns failure? Interruptions, hallucination, and consent/recording need owners before you scale sessions.

At BuildBrain, we treat vendor launches as architecture pressure tests. When OpenAI ships full-duplex voice at a named price, we ask whether your firm has the task map first — inventory the work, cut waste, then decide: AI here, leave that alone — so you are not speeding up the mess.

Talk through your architecture in a free diagnostic — traffic to Ada, a free report, then a paid audit only if the gaps are real. Start on buildbrain.systems or see the architecture audit path.

Demand a voice AI stack you can measure

OpenAI put GPT-Live-1 in the API: full-duplex listen-and-speak, ~$0.05/min voice layer, Live sessions over WebSocket/WebRTC, with published Speak interruption cuts, Full Duplex Bench gains vs GPT-Realtime-2.1, and Tau3 leadership when paired with GPT-6 Astra.

Demand a voice AI stack you can measure: named conversation tasks, interruption baselines, clear voice-vs-backend split, all-in cost per minute, and owners of failure — not another demo that sounds fluent in a quiet room.

If you want a free diagnostic of where your AI stack has demos but no confirmatory design, start on buildbrain.systems.

Disclaimer: This article summarizes a publicly available OpenAI product announcement and is for general informational purposes only. It does not constitute medical, legal, tax, financial, investment, security, or compliance advice. BuildBrain Systems is not a law firm, accounting firm, or registered investment adviser. Metrics, product features, and quotes cited here reflect sources at the time of writing and may change. Readers should verify current information independently and consult qualified professionals regarding obligations specific to their industry, jurisdiction, and circumstances — including applicable federal, state and local requirements. BuildBrain Systems may have commercial relationships with vendors mentioned; where material, such relationships are disclosed. Nothing in this article is an endorsement of any specific AI product or provider.