What did OpenAI announce on September 11, 2026, about launching the GPT Live 1 full duplex voice model in its API, including how it differs from traditional speech to text–language model–text to speech pipelines, its ability to handle simultaneous listening, speaking, interruptio

OpenAI’s API announcement was dated September 10, 2026—not September 11. It introduced GPT‑Live‑1 as a generally available, full‑duplex voice layer: a model intended to make voice agents more natural by listening and speaking concurrently rather than processing discrete turns. 3

What changed technically

• Traditional voice agents chain speech-to-text, a language/reasoning model, and text-to-speech; each handoff adds delay and can lose timing, context, and conversational rhythm. GPT‑Live‑1 combines listening and speaking in one voice model. 3
• It can react to interruptions and acknowledgements as they occur, while a paired backend handles deeper reasoning or tools in the background. 3
• Developers can delegate to an OpenAI backend such as GPT‑6 Astra or to a third-party model, choosing the reasoning, speed, and cost profile appropriate to each task. 3
• OpenAI highlighted system-prompt controls for tone, pace, and conversational style, plus telephony deployment, background-noise and silence handling, longer-session reliability, transcripts/response text, alphanumeric recognition, and keyword biasing. 3

Pricing and access

• GPT‑Live‑1 voice sessions cost $0.05 per minute, billed per second rather than rounded to full minutes. 2
• Backend-model and tool usage are charged separately, so the total cost is the voice-layer charge plus the selected model’s and tools’ usage charges. 2
• The supplied evidence does not substantiate the claimed managed “Presence” enterprise-access details, nor the specific assertion about expanded accents, dialects, and languages. Insufficient evidence.

Reported benchmarks

• OpenAI reported a 30-percentage-point Full Duplex Bench improvement over GPT‑Realtime‑2.1, with GPT‑Live‑1 paired with GPT‑6 Astra at medium reasoning effort ranking first on Tau3. 3
• The supplied evidence supports the broad benchmark claim, but does not independently display the specific 1.41-to-0.798-second latency figures or the 86.2% versus 45.7% Tau3 pass rates. Those exact numbers should therefore be treated as unverified from the evidence provided.

Early-adopter reports

• Yelp said integrating GPT‑Live‑1 into Yelp Host and Hatch improved turn-taking and accuracy over its prior voice architecture; for reservation and food-order calls, it reported meaningful gains in call-handling rates and callers speaking in fuller, more natural utterances. 3
• Speak reported that the model allowed language learners more thinking time before the tutor responded, reducing interruptions by almost 80% compared with previous turn-based systems. 3
• A healthcare-company co-founder/CTO said replacing a cascaded implementation simplified its codebase by 80% and removed 23,000 lines of code, enabling more natural real-time patient conversations. 3