GPT-Live: Voice Models, Access and Session Costs

The GPT-Live 1.1 system card names two successors: GPT-Live 1.1 replaces GPT-Live 1, and GPT-Live 1.1 Quip replaces GPT-Live 1 mini. It identifies 1.1 as the default voice model for Plus and higher tiers, including Codex and Dots, and Quip as the default for other users.

OpenAI’s October 15 announcement says the update is rolling out in its products and is coming soon to the API. It reports 13 new voices, better accents in several languages and more natural delegation. Those product changes do not establish a callable 1.1 API model ID or a Quip API price. For an API integration, the documented model here is gpt-live-1.

As checked October 18, GPT-Live 1 costs USD $0.05 per minute of voice-session duration, billed per second without rounding to a whole minute. A 90-second session costs $0.075; ten minutes cost $0.50 and 100 minutes cost $5 for the voice layer. These are examples calculated from the published rate.

Creating a WebRTC session initially bills 15 seconds of voice duration, or $0.0125 at this rate. Once the session runs, that amount is credited against its duration charges, rather than added on top. Include initialization attempts when reconciling sessions that never start.

Backend model calls and tools are billed separately at their own rates. A ten-minute session with $0.12 of backend and tool usage would total $0.62. That $0.12 is an assumed workload, not an included allowance. Track the session duration and the backend’s billable usage together when comparing assistants.

The audio calculator’s character-based TTS comparison is useful for rendering a fixed script. Live conversation uses session duration, so characters submitted for narration cannot stand in for a GPT-Live bill.

Use OpenAI’s WebRTC quickstart with a microphone, a browser served over HTTPS or localhost, and a trusted server holding your project API key. Its example connects microphone audio, playback and a data channel to a Responses backend that can search the web. Keep the key on the server.

Choose gpt-live-1 for the voice model and give it a short conversation prompt describing when to delegate. The backend gets the detailed task instructions and tools. Responses delegation uses a supported Responses model; client delegation connects an agent or service you run.

Wait for session.started, ask a question requiring the backend, then add a correction while it works. Check that the final answer reflects the corrected request and the actual tool result. Close the session at the end to collect final usage and release the connection. Your application decides whether interrupted backend work should finish or be cancelled.

Use a live session when the listener needs to interrupt, clarify or request work during the conversation. Start with one task whose result you can check, such as asking for a current fact and then changing the question. Measure whether the assistant follows the correction and completes the task, alongside duration and total charges.

For a finished narration file, choose a text-to-speech workflow and compare the cost of every take. Fish Audio Drama 3 is one such option. For delegated routing or classification, the Decisions API guide explains a different component of a voice-agent workflow.