Agent Horizon

OpenAI has opened its full-duplex voice model, GPT-Live-1, to developers through the API, moving it from a ChatGPT-only feature (introduced in July) into a building block anyone can wire into their own products. The model listens and speaks at the same time rather than taking turns, and it hands off harder reasoning or tool use to a separate backend model such as GPT-6 Astra while the conversation keeps flowing. It’s priced at $0.05 per minute for the voice layer, billed per second, with whatever backend model and tools you pair it with billed separately. You can read OpenAI’s own announcement here.

Why this is a genuinely useful shift, not just a spec bump

Voice AI has spent years stitching together three separate systems: speech-to-text, a language model, then text-to-speech, with all the latency and awkward pauses that chaining implies. Collapsing that into one model that processes incoming and outgoing audio simultaneously is a real architectural change, not a marketing repackage, and it shows up in the numbers OpenAI published: a 30-point jump on its Full Duplex Bench versus the prior GPT-Realtime-2.1 model, and a reported first-place finish on Tau3, a benchmark for voice-agent task completion, when paired with GPT-6 Astra. Early partners are backing this up with concrete, if self-reported, results: the language-learning app Speak says it cut unwanted interruptions during learners’ thinking pauses by nearly 80%, and Yelp says the model improved call handling for its reservation and ordering bots.

The caveat worth flagging clearly: the headline price is not the whole bill. GPT-Live-1 is explicitly a voice front-end — it doesn’t do the actual reasoning, retrieval, or business logic on its own. That work goes to a backend model you choose and pay for separately, so a real production voice agent’s cost is the $0.05/minute voice layer plus whatever GPT-6 Astra, Codex, or another model charges for the thinking behind it. That’s a sensible way to separate concerns technically, but it means the eye-catching per-minute price understates what a working phone agent will actually cost, especially since billing runs on talk time (including silence and backend wait time) rather than on tasks resolved.

Competitively, this puts pressure on the standalone real-time-voice vendors (think ElevenLabs or Cartesia-style stacks) that developers previously had to assemble by hand, since OpenAI is now offering the “listen while speaking” capability as a single managed model with telephony support built in. It’s an incremental-feeling but practically significant move: less a headline-grabbing new model and more a plumbing upgrade that should make voice agents noticeably less clunky for anyone building customer support, scheduling, or tutoring tools.