Five cents a minute is the number OpenAI attached to GPT-Live-1 when it opened the model to developers on September 10. That buys the conversational layer of an agent, the part that listens and talks at the same time.
Full duplex is the technical distinction. Audio travels both ways at once rather than waiting for the caller to stop, and the model passes anything needing real thought to whichever model or tool it is paired with. OpenAI’s own example shows a session handing context to Codex and reading the reply back into the conversation.
Benchmarks in the release put GPT-Live-1 ahead of GPT-Realtime-2.1 by 30 points on Full Duplex Bench. On Tau3, which scores voice agents on complete service tasks, it took first place when paired with GPT-6 Astra at medium reasoning effort, and its reported Pass@1 rate on airline, retail and telecom work reached 86.2 percent against 45.7 percent for the older model.
What developers receive is a list of practical features: background noise handling, retention across long calls, telephony, speech recognition transcripts, and turn detection for teams that still want explicit boundaries.
Cheap conversational voice changes what gets built. Support lines, kiosks and reservation desks become viable products rather than experiments once the cost of an agent talking drops below the cost of playing a hold message. Regulators are circling the same ground, asking how callers learn that the voice on the line is software.
