What changed
OpenAI launched GPT-Live-1 in the API on September 10, 2026. The model handles incoming and outgoing audio together rather than relying on a conventional speech-to-text, text-model and text-to-speech chain. It can listen while speaking, react to interruptions and acknowledgements, manage silence and background speech, emit transcripts and response text, and support explicit turn detection. Developers can pair the voice layer with a separate backend model or agent harness for reasoning and tool use. OpenAI prices the front-end voice layer at $0.05 per minute and supports browser/server/telephony deployment paths.
Why it matters
Voice-agent engineering has often been an orchestration problem: every transition between transcription, reasoning and synthesis adds latency and creates another place where interruptions, timing or context can go wrong. GPT-Live-1 collapses the conversational front end into one model while keeping expensive reasoning separable. That gives builders a new cost and architecture choice: pay a predictable per-minute voice charge, then route simple or complex work to whatever backend model fits the task. It also makes interruption handling and telephony behavior a model capability rather than entirely application code.
The voice loop becomes one model
GPT-Live-1 jointly processes incoming and outgoing audio, allowing it to react while it is speaking. OpenAI positions this as an alternative to cascaded STT–LLM–TTS systems where each handoff adds latency and brittle state transitions.
Reasoning can stay behind the voice layer
The model can delegate deeper reasoning and tool calls to a separate backend model, including third-party models. Developers therefore do not have to use the same model for natural conversation and expensive planning or tool execution.
Interruption handling is a core capability
GPT-Live-1 is designed to handle pauses, backchannels, self-corrections and users speaking over the agent. It also provides native transcripts, response text, keyword biasing and turn detection for applications that still need explicit conversational boundaries.
Telephony is part of the launch surface
OpenAI explicitly supports phone-call use cases as well as browser and server integrations. That makes the release relevant to customer support, reservations, scheduling and other workflows where full-duplex timing matters more than raw text-model benchmark scores.
Voice cost is separate from reasoning cost
OpenAI lists $0.05 per minute for GPT-Live-1's front-end voice layer. The backend model and tools are additional choices and costs, so teams should model end-to-end cost per successful call rather than treating five cents per minute as the complete agent price.