Enterprise

OpenAI opens GPT-Live-1 to developers at $0.05 a minute, putting phone agents within reach of small teams

The full-duplex voice model behind ChatGPT Voice is now callable from the API, collapsing the old speech-to-text-to-speech pipeline into a single billable minute.

By Asher Glick· Enterprise software desk·2026-09-15·3 min read

Photo: Unsplash / Berkeley Communications — Customer service agent wearing a headset at a desk

OpenAI released GPT-Live-1 to developers on 10 September 2026 at $0.05 per minute, exposing the full-duplex voice model that already powers roughly 150 million weekly ChatGPT Voice users through a new v1/live/sessions endpoint. On the same day, the Agents API entered public beta. The pricing is the story: a phone agent that answers reservations, qualifies inbound leads, and holds an interruption-aware conversation is now a line item, not an engineering project.

The architecture matters because it retires one. Cascaded stacks (speech-to-text into an LLM into text-to-speech) posted P50 latencies of roughly 800 milliseconds to two seconds, and telephone deployments piled on another 80 to 150 milliseconds of carrier round-trip. GPT-Live-1 answers at 0.798 seconds on OpenAI’s Full Duplex Bench v1, about 43% sooner than GPT-Realtime-2.1’s 1.41 seconds. Twelve voices ship at launch: Quartz, Ripple, Vesper, Willow, Stone, Gleam, Meridian, Bossa, Tempo, Beacon, Delta, and Cinder.

Early operators are describing what natural pacing buys them. Yelp, deploying through its Yelp Host AI reservation system, reports better call-handling and callers speaking in fuller sentences, according to CTO Alex Levy. Andrew Hsu, co-founder and CTO of Speak, credits GPT-Live-1 with cutting false interruptions by almost 80% against the platform’s previous turn-based stack. An unnamed healthcare customer told OpenAI it deleted 23,000 lines of glue code.

The five-cent headline hides a meter. OpenAI’s cost guide bills active session time whenever the user speaks, the assistant speaks, both are silent, or the backend is working. 1,000 hours of calls costs $3,000 before model tokens, and WebRTC session initialisation bills 15 seconds. Backend choice does the real damage: GPT-5.6 Luna runs about $1.20 per million output tokens; GPT-6 Astra runs $50. The migration guide warns bluntly that “interrupting speech does not automatically cancel backend work.”

Other limits shape scope. Knowledge cuts off 31 July 2025. No image or video input. Concurrency runs 25 sessions at Tier 1 to 500 at Tier 5. GPT-Live-1 sessions qualify for Zero Data Retention and European processing; the Agents API does neither and is US-only.

For a team of five to thirty, the disciplined first deployment is a scoped inbound line (bookings, qualification, after-hours triage) rather than a wholesale replacement of the human queue. The infrastructure finally isn’t the hard part. Scoping is.

Sources