Skip to content
Neural Daily

ModelsNews

GPT-Live-1.1 hits the API with cheaper standby mode

🇭🇺 Olvasd el magyarul →

OpenAI has made GPT-Live-1.1, the full-duplex voice model behind ChatGPT voice since October, available in the API. The new standby mode drops the price of a silent session to $0.02 per minute, down from $0.05 when active.

· 2 min read

Az OpenAI Developers GPT-Live-1.1 modelloldala: full-duplex hangmodell, aktív ár 0,05 dollár, standby 0,02 dollár percenként
Image: OpenAI · Screenshot · Source

What launched?

OpenAI announced on its developer forum that GPT-Live-1.1 is available in the API (OpenAI Community). It has powered voice in ChatGPT Plus and Pro since October 15. It is full-duplex, so it listens while it speaks, and it hands complex work to a backend model of your choice without breaking the conversation.

Compared with GPT-Live-1, OpenAI says it transcribes speech more accurately, keeps a more consistent accent and voice across languages, handles letters and numbers better, and follows instructions and completes tasks more reliably (OpenAI Developers).

Pricing and standby

  • Active session: $0.05 per minute.
  • Standby: $0.02 per minute. Backend model and tool usage is billed separately.
  • Standby works for GPT-Live-1 and 1.1 WebRTC sessions: set session.wake_policy.type to speech and session.idle_timeout_s to a whole number from 5 to 60.
  • After the set silence, inference pauses while the connection and session ID stay the same; when someone speaks, the session resumes with its history (
  • [object Object]
  • ).
  • The first response to a short utterance after standby may take longer.

An example: if 4 minutes of a 10-minute call are silent, the voice session costs 6 × $0.05 + 4 × $0.02 = $0.38 with standby, versus $0.50 without. For support or hold-heavy calls, that adds up fast.

Watch out for

  • It only runs on the v1/live/sessions endpoint (WebRTC or WebSocket). Chat Completions, Responses and the Realtime API are not supported.
  • GPT-Live-1.1 and GPT-Live-1 have separate concurrent-session limits.
  • Instructions can be up to 16,384 tokens, and the voice (marin by default) can only be changed by starting a new session.

Why it matters

For voice agents, continuous inference has been the main cost, and standby targets exactly that. The full-duplex, delegating design also means the voice model does not need to know everything: reasoning and tool calls happen in the backend. Among alternatives, Sophon-Realtime runs on a single GPU (Sophon-Realtime v1) and Microsoft's MAI Realtime 1 starts at $0.01 per minute (Microsoft Ignite 2026). If you are picking a voice stack now, compare against those.

    By Mikola Máté Patrik

    Produced with the help of AI, reviewed by an editor. How we make our articles

    Reviewed and edited by the Neural Daily desk. About

    Related articles