OpenAI Debuts GPT-Live-1 Voice Model for $0.05 a Minute

OpenAI has released GPT-Live-1, a full-duplex conversational voice model now available in its API. Unlike traditional voice systems that chain separate speech-to-text, language, and text-to-speech models, GPT-Live-1 listens and speaks simultaneously. This architecture allows the model to handle interruptions, backchannels, and mid-thought corrections in real time. The service costs $0.05 per minute for the voice layer, which operates on top of the developer's chosen backend reasoning model.

In performance evaluations, GPT-Live-1 outperformed GPT-Realtime-2.1 by 30 percentage points on the Full Duplex Bench. When paired with GPT-6 Astra at medium reasoning effort, it achieved the top ranking on Tau3, a benchmark measuring frontier voice-agent intelligence. Architecturally, the model delegates reasoning and tool calls to a backend text model of the developer's choice, such as GPT-6 Astra, Luna, or even third-party options. This allows developers to keep callers engaged with natural filler while a backend model processes complex tasks.

The release includes telephony support, background noise handling, keyword biasing, and 12 new voices: Quartz, Ripple, Vesper, Willow, Stone, Gleam, Meridian, Bossa, Tempo, Beacon, Delta, and Cinder. Early adopters report massive reductions in system complexity. One healthcare customer eliminated 23,000 lines of pipeline code, shrinking their codebase by 80 percent. Additionally, the language-learning platform Speak used the model to reduce conversational interruptions by nearly 80 percent.

Other early partners deploying the model include Yelp for reservation calls, Cognition for voice interactions with its Devin coding agent, and Fin for customer support. Developers can also access a reference pattern to connect GPT-Live-1 with Codex, enabling users to edit code repositories through spoken conversation. By maintaining backend flexibility while resolving the latency and complexity of cascaded pipelines, the model fundamentally shifts the economics and architecture of voice-agent development.