Build more natural voice experiences with GPT‑Live‑1 in the API
Sep 9, 2026, 5:00 PM · OpenAI
OpenAI’s GPT‑Live‑1 hits the API as a full-duplex voice front end — listen and speak at once, hand hard reasoning to a backend model, priced at five cents a minute.
Why it matters
OpenAI is launching GPT‑Live‑1 in the API, bringing ChatGPT’s natural, full-duplex voice conversations to developers. The model listens and speaks at the same time, handles interruptions in one stack, and can delegate deeper reasoning and tool calls to a backend text model such as GPT‑6 Astra or a third-party model.
Early customer notes are concrete. Speak says GPT‑Live‑1 cut interruptions during thinking pauses by almost 80% versus previous turn-based systems. OpenAI reports a 30 percentage-point gain on Full Duplex Bench over GPT‑Realtime‑2.1, and says pairing Live with GPT‑6 Astra at medium reasoning effort ranks #1 on Tau3 for end-to-end spoken agent tasks. Pricing for the front-end voice layer is $0.05 per minute.
From the desk
We’re watching voice agents leave the STT–LLM–TTS Rube Goldberg machine. Chained pipelines add latency and break when someone talks over the bot. GPT‑Live‑1 reasons over incoming and outgoing audio together, keeps tone and pace steerable from the system prompt, and is meant to handle background noise and silence without narrating every step. Telephony support is explicit — reservations, support lines, the phone call as a product surface.
Customer quotes on the page sketch the architecture win. A co-founder at one health-facing team says Live simplified their codebase by 80% and removed 23K lines versus a cascaded build. Yelp’s CTO says Host and Hatch saw better turn-taking and accuracy, with callers speaking fuller sentences. Intercom’s Fin team frames full-duplex as the shift from stop-start rhythm toward a real phone call. Cognition’s Walden Yan ties Live to talking with Devin like a teammate. We’re treating those as vendor-selected testimonials, not independent audits.
Our read: this is useful AI for anyone still gluing three models into a call center. Separating a cheap, interruptible voice layer from a selectable backend (Astra for hard cases, something lighter for scheduling) is the right cost shape. The downside if it scales is voice that feels human enough to skip disclosure, plus new failure modes when delegation lags and the caller thinks the agent “heard” a commitment it hasn’t executed. I’m watching alphanumeric accuracy on account numbers, long-session drift, and whether $0.05/min plus backend tokens still beats a tuned cascade once volume spikes.
Context
GPT‑Live‑1 natively provides ASR transcripts and response text, supports keyword biasing, and offers turn detection even though it is not a turn-based model. OpenAI is expanding the real-time voice catalog across accents, dialects, and languages. Presence is positioned as an enterprise layer on top of Live for trusted agents that can escalate to people. Codex connection examples show Live delegating spoken questions into a repo-aware coding thread.
Who feels it
- Voice-agent developers
- A single full-duplex model can delete a lot of glue code. Keep an explicit handoff contract when Live delegates tools so callers aren’t promised work that hasn’t finished.
- Contact centers
- Telephony support plus interruption handling is the product. Measure contain rates and escalation quality, not just demo naturalness.
- Language-learning and consumer apps
- Speak’s nearly 80% interruption cut is the UX signal. Silence tolerance may matter as much as speech quality.
What to watch
- Independent Full Duplex Bench and Tau3 replications outside OpenAI’s eval framing.
- Real per-minute cost once backend Astra or third-party reasoning is added at production concurrency.
- How fast new voices and languages ship, and whether custom voice access stays sales-gated.