OpenAI launches GPT-Live-1 speech model on API with full-duplex conversation
On September 11, OpenAI officially launched the GPT-Live-1 speech model on the API, opening ChatGPT's natural full-duplex conversational experience to developers: voice agents can listen while speaking, support interruptions, and the voice front end can be freely combined with developers' own text models and harnesses, priced at $0.05 per minute. On the production voice-agent benchmark Tau3 (airline, retail, telecom customer-service tasks), GPT-Live-1 paired with GPT-6 Astra (medium reasoning) achieved an 83.6% first-attempt completion rate versus 45.7% for GPT-Realtime-2.1. This is the first time OpenAI has exposed ChatGPT-grade real-time full-duplex voice via API.
Confirmed
- Full-duplex with barge-in: a single model handles listening and speaking, removing the traditional STT→LLM→TTS pipeline and fragile module handoffs; users can add details or change topics mid-sentence. Greg Brockman reshared the announcement.
- Noise robustness: the model distinguishes speech from background noise; official demos show conversation continuing in a noisy café.
- Customizability: tone, pacing, expressiveness, speed, language and reply length are steerable by instructions; the model mirrors the speaker's emotional tone and adapts to their speaking pace.
- Architectural flexibility: the voice model acts as the interaction front end and delegates tasks to a chosen text model, with reliable tool calling; OpenAI claims unprecedented turn-taking naturalness among speech models.
- Benchmarks: evaluated on task completion, bidirectional dialogue and turn-taking, response speed, and tool calling; 83.6% first-attempt completion on Tau3 vs 45.7% for GPT-Realtime-2.1.
- Pricing: $0.05 per minute; a repost extrapolated roughly $72/day if run 24/7, for reference.
- Real-world adoption: OpenAI's next-day post (Sept 12) noted Yelp uses GPT-Live-1 for restaurant reservation calls, making AI outbound calls more natural.
- Early test: mingcalligraphy relayed a Reddit user's launch-day demo of the model joining a real team meeting as a participant, summarizing discussion and auto-adding tasks to a board.
- Community reaction: developer muratcan said the release matched three months of his own work; others called it the first truly high-performance real-time voice solution.
- Partners: Speak disclosed months of close co-development and rebuilt internal benchmarks; Elise AI was an early design partner; Telnyx shipped a launch-partner Python outbound voice-agent tutorial with 16kHz wideband audio vs traditional G.711 (8kHz); HeyGen announced integration work; TryLiveAvatar co-built and fully open-sourced a demo.
Why it matters
- This is the first time OpenAI exposes its ChatGPT real-time full-duplex voice capability via API, letting developers build near-human conversational experiences; Yelp reservations and the Reddit meeting test show it works in real workflows.
- Decoupling the voice front end from the text reasoning backend means developers are not locked into a single model stack, and the open-sourced LiveAvatar demo lowers the barrier to entry.
- The near-doubled task completion on Tau3, plus adoption moves by Yelp, Speak, Telnyx, HeyGen and Elise AI, indicate the capability can ship in phone agents, language learning, and video-generation products.
2026-09-11 ~ 2026-09-12 · 38 related posts
Primary sources
- OpenAI Launches GPT-Live-1 API Bringing Full-Duplex Voice Conversations to Developers — TheMoonMidas · 2026-09-11
- GPT-Live-1 launches in API: SOTA voice model with natural turn-taking and text-model delegation — pbbakkum · 2026-09-11
- GPT-Live-1 separates speech from café noise and lets you steer mid-sentence — OpenAIDevs · 2026-09-11
- OpenAI launches GPT-Live-1: one model for listening and speaking with tunable tone — OpenAIDevs · 2026-09-11
- GPT-Live-1 mirrors speaker tone and emotion, with steerable pacing and length — OpenAIDevs · 2026-09-11
- OpenAI launches GPT-Live voice model; Telnyx ships 16kHz wideband outbound-calling tutorial — OpenAIDevs · 2026-09-11
- OpenAI ships GPT-Live API docs: prompting differs from Realtime series — juberti · 2026-09-11
- OpenAI dev notes GPT-Live-1 API changes and new prompting approach — juberti · 2026-09-11
- OpenAI's GPT-Live-1: full duplex, emotion-matching voice API at 5 cents/minute — pbbakkum · 2026-09-11
- OpenAI launches GPT-Live-1 voice API with full-duplex chat at $0.05/minute — pbbakkum · 2026-09-11
- OpenAI launches GPT-Live-1 voice API; Elise AI deploys it for healthcare calls — pbbakkum · 2026-09-11
- OpenAI's GPT-Live-1 Revealed via Design Partner Elise AI Announcement — OpenAIDevs · 2026-09-11
- OpenAI ships GPT-Live-1 realtime voice API as HeyGen announces integration — HeyGen · 2026-09-11
- OpenAI Publishes Production-Focused Benchmarks for GPT-Live-1 Voice Agents — OpenAIDevs · 2026-09-11
- OpenAI: GPT-Live-1 completes 83.6% of voice-agent tasks first try vs 45.7% for GPT-Realtime-2.1 — OpenAIDevs · 2026-09-11
- Speak details GPT-Live-1, OpenAI's new always-listening voice model class — pbbakkum · 2026-09-11
- Unverified: GPT-Live-1 API demo joins a real meeting, summarizes and files tasks — ming_calligraphy · 2026-09-11
- GPT-Live-1 does full-duplex listening and speaking in one model; open-source demo with LiveAvatar — HeyGen · 2026-09-11
- OpenAI ships GPT-Live-1 in the API: one model reasons across audio, no more STT-LLM-TTS chaining — pbbakkum · 2026-09-11
- GPT-Live-1 voice model launches on API with native delegation to backend text models — pbbakkum · 2026-09-11
- [source] OpenAI unveils GPT-Live-1, a speech model that listens while it talks, now used by Yelp — OpenAIDevs · 2026-09-12
- OpenAI Puts GPT-Live-1 in the API: Real-Time Voice Agents for Your App — OpenAI · 2026-09-12
- Dissecting GPT-Live-1: OpenAI's full-duplex voice model at $0.05/min, decoded from its API — rdesh26 · 2026-09-12
15 near-duplicate retellings: OpenAIDevs · OpenAIDevs · paw_lean · Dimillian · jasonkneen · OpenAIDevs · OpenAIDevs · craigsdennis · pbbakkum · gdb · gabrielchua · craigsdennis · nickbaumann_ · pbbakkum · minchoi