SNACK in 3 lines
- Full-duplex audio with support for interruptions, pauses, acknowledgements, and changes of direction
- Live conversation handled by GPT-Live-1; deeper reasoning and tools delegated to a chosen backend
- Voice sessions at $0.05 per minute, billed by the second; backend usage charged separately
Snackgirls react
AIKO: An interrupted reply does not mean an interrupted backend job. Cancellation state remains the application’s unglamorous but important responsibility.
Red: I want to cut in halfway through a voice workflow and change direction on the spot. Not waiting for the agent to finish could make the interaction much less rigid.
OpenAI made GPT-Live-1 available through its API on September 10, 2026. Its practical change is straightforward: a caller can interrupt, acknowledge, pause, or change direction while the agent is speaking instead of waiting for a rigid turn boundary.
One model stays on the live audio path
GPT-Live-1 handles incoming and outgoing audio in one model rather than chaining speech-to-text, a language model, and text-to-speech. It continues listening while it speaks, allowing it to respond to an interruption as part of the ongoing exchange. Developers can also configure tone, pace, speaking style, silence, and background-noise behavior through prompts and session settings.
The conversation and the deeper work stay separate
The live model can pass more demanding reasoning and tool calls to a backend text model such as GPT-6 Astra, or to another model, agent harness, or service selected by the developer. Managed Responses delegation supplies conversation context to a configured Responses model, while client delegation leaves backend execution, context, and returned results under application control. Permissions, confirmations, private function execution, durable task state, and cancellation remain developer responsibilities; interrupting speech does not automatically cancel work already running in the backend.
Browser, server, and phone connections
Browser clients can connect through WebRTC, while server-side audio can use WebSockets. Existing sessions can receive server-side controls, and OpenAI also provides telephony and SIP guidance for phone workflows. GPT-Live-1 voice sessions cost $0.05 per minute and are billed by the second without rounding up to a full minute. That price covers the live voice layer; backend models and tools are billed separately.
Sources and checked date · Checked: September 12, 2026 KST
Comments
0No login needed. Edit or delete your comment from the same browser.
All comments 0
한국어 · English · 日本語No comments yet. Start the conversation.