GPT-Live-1
OpenAI's full-duplex voice layer — listens and speaks simultaneously, with reasoning delegated to a backend model.
About
GPT-Live-1 is OpenAI's full-duplex voice model: it listens and generates speech at the same time instead of taking turns, so voice agents handle interruptions, pauses, backchannels, and background noise the way a phone call does. It first rolled out inside ChatGPT Voice on July 8, 2026 and reached the API on September 10, 2026.
The key architectural split: GPT-Live-1 is the voice layer only. Reasoning and tool use are delegated to a backend model of the developer's choice (GPT-6 Astra, Luna, Terra, or third-party models), keeping the voice loop fast while heavy thinking runs in the background. Interrupting the speech does not cancel the backend task. It ships with 12 voices across accents and languages, telephony support, per-second billing, native transcripts, keyword biasing, and turn detection.
On OpenAI's own Full Duplex Bench it scores 30 percentage points above GPT-Realtime-2.1, with turn-taking latency down to 0.8s from 1.4s and tool-calling accuracy up to 87% from 60% in a banking voice-support test pass rate rising from 12.4% to 32%.
Specs
Context window
Pricing
Free quota
API access
Weights
Released
Benchmarks
Provider-reported (OpenAI): Full Duplex Bench +30pp over GPT-Realtime-2.1; turn-taking latency 0.8s vs 1.4s; tool-calling accuracy 87% vs 60%. Independent (Artificial Analysis): 83.2 on the Speech-to-Speech Index as GPT-Live-1 (Astra, Medium), behind Grok Voice 3.0 at 86.9.
Provider-reported unless a source says otherwise. Benchmarks are marketing-adjacent by nature — treat them as directional.
Use cases
Trust score — 7/10
Release facts (July 8, 2026 ChatGPT rollout; September 10, 2026 API; $0.05/min per-second billing; voice-layer/backend split; 12 voices; telephony; AA index 83.2) corroborated across OpenAI's write-up, DataNorth, AI Tech Connect, and Unite.AI. Docked for: (1) the headline improvement figures are OpenAI's own evaluations; (2) two-line billing (voice layer plus backend) makes real costs harder to estimate; (3) proprietary with no weights.
How the trust score is built
Full methodology →- Adoption · stars/downloads, log-scaled vs peers
- 7.0/10
- Maintenance · recency, releases, issue load
- 10.0/10
- Openness · license + docs
- 2.0/10
- Manual review · sanity read + red-flag check
- 8.0/10
- · Adoption 7: powers ChatGPT Voice for paid users; API since Sep 10 2026 with /v1/live/sessions
- · Maintenance 10: current OpenAI voice stack; iterated via Sep 9 voice update
- · Openness 2: hosted API only, no weights
- · Manual 8: release facts (Jul 8 2026 ChatGPT rollout; Sep 10 2026 API; $0.05/min per-second billing; voice/backend split; 12 voices; AA 83.2) corroborated across OpenAI write-up, DataNorth, Unite.AI; docked for provider-only improvement figures, two-line billing opacity, closed weights