Voice AI just got cheap enough to build with

For years, building a voice assistant meant linking three separate systems: one to turn speech into text, one to reason, and one to speak back. Every link added delay, and whenever a caller interrupted, the whole chain broke. OpenAI's GPT-Live-1, now open via API, collapses those three steps into one model.

The price is the news. At $0.05 per minute, a five-minute call costs 25 cents. That is low enough for a small startup to run real phone agents without a big budget. One company that moved to GPT-Live-1 reportedly deleted 23,000 lines of code that existed only to handle interruptions — the model deals with that on its own now.

GPT-Live-1 is what engineers call full-duplex — it listens and speaks at the same time, just like a person does. Callers can cut in mid-sentence without the AI losing its place. In testing, it outperformed the previous best voice model by 30 points on the industry benchmark. That gap matters most in anything customer-facing, where a broken interruption or an awkward pause kills trust.

For anyone building software that talks to people — support lines, intake forms, booking tools, sales calls — this shifts what a small team can actually ship. The one thing to know: $0.05 covers only the voice layer; the AI reasoning behind it is billed separately, so total cost depends on which model you connect. But for most teams, this makes a real voice agent possible for the first time.