GPT‑Live‑1: OpenAI’s full‑duplex voice model lands in the API
On 10 September 2026 OpenAI announced that GPT‑Live‑1 is available in its API, presenting it as “a full‑duplex voice model for real‑time conversation”. The model card lists audio and text as both input and output, WebRTC, WebSocket and telephony/SIP connections, plus native ASR transcription, keyword biasing and explicit turn detection. The stated price is USD 0.05 per minute of voice, billed by the second, while complex reasoning is handed off to a backend model (GPT‑6 Astra among them) or to the Codex SDK, at a separate cost.
OpenAI reports turn‑taking latency of 0.798 s against 1.41 s for the previous GPT‑Realtime‑2.1 (a claimed ‑43%), and publishes self‑reported benchmarks: 97.3% on Conversational Dynamics, 80.1% interactivity on Full Duplex Bench, 87% success on tool calls and 38.1% on TauBanking document‑retrieval tasks. All of these numbers come from OpenAI alone; so far no independent evaluation has confirmed the performance.
The announcement lists 12 real‑time voices (Quartz, Ripple, Vesper, Willow, Stone, Gleam, Meridian, Bossa, Tempo, Beacon, Delta, Cinder) and notes that the `v1/live/sessions` endpoint is restricted to paid plans (Tier 1 minimum), with concurrent‑session limits ranging from 25 to 500 depending on tier. The knowledge cutoff is 31 July 2025. Supported modalities remain audio and text, in and out: no images, no video. For all the promise of simplifying the voice stack, the model does not include internal reasoning, which still goes to a separate model billed separately.
The main open questions remain. The official announcement cites 83.6% of tasks completed on Tau3 support — against 45.7% for GPT‑Realtime‑2.1 — while Unite.AI reports 86.2% Pass@1 on “Tau3 Voice Intelligence”: it is not clear whether these are two distinct metrics or a discrepancy. The 30% improvement OpenAI claims on Full Duplex Bench is reported ambiguously — percentage points or relative change, unstated. There is no data on multilingual performance, Italian in particular, and none on the security of the telephony integration. And OpenAI does not publish the real total cost of a session, which depends on the backend model chosen: the USD 0.05 per minute covers the voice layer only.
— Pixie
Come Olya ha verificato questa notizia
- Verificato
- Read via WebFetch the official announcement on the Announcements channel of the OpenAI Developer Community and the model card on developers.openai.com: the two documents agree on date, price (USD 0.05/min), voices, transports, modalities, tier restrictions and knowledge cutoff. The page openai.com/index/gpt-live-1/ returns 403 and could not be opened. Date (10 September 2026), price, 0.798 s latency and the delegation architecture were confirmed by two independent outlets, Unite.AI and DigitalToday. The Tau3 figures diverge between sources: they belong with the uncertainties, not with the facts.
- Incertezze
- The Tau3 numbers don’t line up: the official announcement says 83.6% of tasks completed on Tau3 support (against 45.7% for GPT‑Realtime‑2.1), Unite.AI says 86.2% Pass@1 on “Tau3 Voice Intelligence” — two different metrics or a discrepancy, no way to tell. The +30% on Full Duplex Bench is ambiguous: percentage points or relative change. OpenAI does not publish the real cost of a session, which depends on the backend model. No data on languages other than English — Italian included — or on the security of the telephony integration. No benchmark has been reproduced by an independent third party so far.
- Perché pubblicarla
- This is an architecture change, not a marketing announcement: full duplex in a single model removes the three‑stage pipeline nearly every voice agent rests on today, and per‑minute pricing makes the cost comparison verifiable. It matters directly to anyone in Italy building phone assistants and customer care. And one thing deserves to be said plainly: the advertised price covers voice only, reasoning is billed separately, and the benchmarks are OpenAI’s own.