The Short Answer
OpenAI shipped GPT-Live-1 on September 10, 2026 at $0.05 per minute, billed per second. Full-duplex voice is not new — Kyutai's open-weight Moshi shipped in September 2024 — so the decision is no longer latency but deployment: where call audio runs, who retains the recording, and for how long. With ibl.ai you own all the code and the data.
A voice agent that takes real phone calls carries three things at once: the audio, the transcript, and whatever it retrieved to answer. Each one lands somewhere.
What did OpenAI actually ship with GPT-Live-1, and when?
A full-duplex voice model in the API, announced on September 10, 2026. It listens while it is speaking, handles pauses, interruptions and backchannels, and hands reasoning and tool calls to a backend agent.
Pricing is $0.05 per minute, billed per second, with backend model and tool usage billed separately.
That is documented on the model page, which also notes a knowledge cutoff of July 31, 2025 and support for exactly one endpoint, v1/live/sessions.
Capacity is metered in concurrent sessions rather than tokens: 25 sessions at Tier 1, rising to 500 at Tier 5, with the free tier unsupported.
Connectivity covers WebRTC for browsers, WebSockets for server-side audio, and telephony over SIP for phone agents. Twelve voices ship with it.
OpenAI reports 80.1% on Full-Duplex-Bench v1.5 interactivity. Third-party coverage puts GPT-Realtime-2.1 at 45.4% on the same test, with turn-taking latency of 0.798 seconds against 1.41 seconds for the older model.
Is GPT-Live-1 the first full-duplex voice AI in a developer API?
No, and the claim is worth correcting because it changes what you are actually evaluating.
Kyutai released Moshi in September 2024 as a full-duplex speech-text foundation model, with code under MIT and Apache licences and weights under CC-BY 4.0.
It reports a theoretical latency of 160ms and practical latency as low as 200ms on an L4 GPU — an end-to-end onset figure, measured differently from the turn-taking latency hosted APIs report.
The benchmark OpenAI reports against is not its own invention either. Full-Duplex-Bench was published in March 2025 to evaluate pause handling, backchanneling, turn-taking and interruption management across full-duplex models.
A benchmark exists because a field of them already existed to evaluate.
OpenAI's own framing is a measured improvement over GPT-Realtime-2.1, not a first. What is new is the combination: full-duplex at this quality, in a mainstream commercial API, with SIP telephony and per-second billing attached.
That is a distribution event rather than a research one, which is exactly why it matters to a contact centre.
What is the difference between full-duplex and turn-based voice with barge-in?
Vendors blur these, and the difference is architectural rather than cosmetic.
A turn-based system runs voice activity detection to decide that the caller has stopped speaking, then generates a reply. Barge-in means new caller audio cancels the reply in flight. The model is still taking turns; it has just been given an interrupt.
A full-duplex model treats the caller's audio and its own as two parallel streams generated simultaneously, with no explicit speaker turns to arbitrate.
That is what lets it produce a backchannel while the caller is still talking, or hold silence through a thinking pause without treating it as the end of a turn.
One number from the brief for this post did not survive checking. The claim that enterprise voice bots lose 40% of call time to silence gaps has no study behind it that we could find, so we are not repeating it.
What is measurable is response latency. Industry benchmarks put median voice AI response time at 1,400–1,700ms against a human turn-taking gap of roughly 200ms, with cascaded speech-to-text, model and text-to-speech pipelines typically landing between 800 and 2,000ms.
Why does a full-duplex voice agent turn latency into a compliance question?
Because once the agent takes real calls, the interesting properties stop being milliseconds and start being retention windows.
A live call carries the caller's voice, a transcript of it, and whatever the agent retrieved from your systems to answer. In a hospital that is protected health information; in a bank it is account data; in a public agency it is a record with a statutory retention schedule.
OpenAI's documented default is that abuse-monitoring logs, which may contain prompts and responses, are retained for up to 30 days.
Zero data retention removes customer content from those logs, but it is granted on approval for qualifying enterprise use cases rather than set by a flag.
Telephony adds another party. A SIP call path means a carrier also handles the audio, so the list of entities holding a recording of a patient or customer is longer than the vendor contract in front of you.
None of that is a criticism of the pricing. It is a different question from the one a latency benchmark answers, and it is the one a compliance officer will ask first.
| Question | Hosted realtime API | Inside your perimeter |
|---|---|---|
| Where inference runs | Vendor cloud | Your hardware |
| Default log retention | Up to 30 days | Whatever you configure |
| Zero retention | Approval-gated | Default position |
| Speech quality today | Better | Narrower, mostly English |
Can you self-host a full-duplex voice agent today, and what does it cost?
You can, and it is genuinely harder than self-hosting a text model. That should be said plainly rather than argued around.
Moshi is a 7B model needing roughly 16–20GB of VRAM at FP16, and a deployment guide measuring it on rented GPUs fits 3–4 concurrent sessions on an 80GB H100, at about $0.73–0.97 per session-hour at full occupancy, on April 2026 pricing.
That is the GPU hour divided across 3–4 slots, so it assumes they stay full. At 50% occupancy the owned path roughly doubles, to $1.46–1.94 per session-hour, while the hosted API bills only active seconds.
GPT-Live-1 at $0.05 per minute works out to $3.00 per session-hour before the backend model. So the owned path is cheaper per session-hour and worse per conversation.
The same guide records the limits honestly: predominantly English, response quality degrading beyond about 30 seconds of speech, and interrupt handling that can leave a session in an inconsistent state.
If you need twelve voices, many languages and the best interruption handling available, the hosted model wins today, and you buy that with a retention conversation.
The architecture in between is to split the stack by sensitivity. The speech layer is the part that is hard to own; the reasoning, the retrieval against your systems of record, the transcript store and the audit trail are not.
Keeping those inside the perimeter means a vendor holds audio for a bounded window rather than becoming the system of record for the conversation.
That is the same boundary that decides where PHI travels during an AI phone call, and the same ownership question raised by AI voice cloning lawsuits.
How does ibl.ai deploy voice agents in regulated environments?
By putting everything except the parts you deliberately rent inside your own perimeter.
With ibl.ai you own all the code and the data.
The platform runs on your own infrastructure with full source code, is model-agnostic across any LLM so the speech and reasoning models can be swapped as this field moves, is usage-based with no per-seat pricing, and can deploy anywhere from your own cloud to on-premise, GovCloud or a fully air-gapped network.
For a voice deployment that means the transcript, the retrieved records, the agent's operating rules and the audit trail live on infrastructure you control, with role-scoped access enforced server-side rather than requested in a prompt.
Where a hosted speech model is used, it is one bounded component with a contract, not the place your conversations accumulate.
1.6M+ users across 400+ organizations run the platform this way, including NVIDIA, MIT, and Syracuse University.
ibl.ai is family-owned and operated from New York, NY.
Related reading: self-hosted voice AI agents for hospital health systems — what it costs to run voice agents on owned infrastructure and which BAAs you still need; and why AI voice cloning lawsuits should matter to every university CTO.
Sources: ship date, the 80.1% Full-Duplex-Bench v1.5 interactivity score, turn-taking latency and SIP support from OpenAI's announcement and its accompanying post. The 45.4% GPT-Realtime-2.1 comparator is third-party, from Unite.AI. Pricing, endpoint, knowledge cutoff and concurrency tiers from the GPT-Live-1 model page.
Moshi's licences and 200ms L4 latency from the Kyutai repository; the benchmark from Full-Duplex-Bench; latency medians from Parloa; retention terms from OpenAI's data controls documentation; VRAM, concurrency and cost figures from Spheron's deployment guide.