OpenAI ships GPT‑Live‑1 in the API — full-duplex voice at $0.05/min

· opgehaald 12:07

GPT‑Live‑1 (10 Sep) brings ChatGPT’s full-duplex voice to the API: listen+speak together, interrupt naturally, delegate reasoning/tools to Astra/Terra/Codex. +30pp Full Duplex Bench vs Realtime‑2.1; Speak saw ~80% fewer tutor interruptions. Telephony + more voices; $0.05/min front-end.

On 10 Sep 2026 OpenAI launched GPT‑Live‑1 in the API — the full-duplex voice model first seen in ChatGPT Voice. A single model listens and speaks at once (no brittle STT–LLM–TTS cascade), with smooth interruption handling, tone/pace/style via system prompt, silent context + background-noise handling, long-session reliability, and telephony for phone agents. It can delegate deeper reasoning and tool calls to a backend text model (e.g. GPT‑6 Astra, Terra/Luna) or Codex; natively returns ASR transcripts and response text, with turn detection when apps still want turn boundaries. Evals: +30 percentage points on Full Duplex Bench vs GPT‑Realtime‑2.1; paired with Astra (medium) it ranks #1 on Tau3. Speak cut tutor interruptions during thinking pauses by almost 80% vs turn-based systems; Yelp, Intercom Fin, and Cognition cite clearer turn-taking. Pricing: $0.05 per minute for the front-end voice layer (backend billed separately). Docs: model gpt-live-1, Live sessions endpoint; also powers OpenAI Presence for enterprise voice agents.