Back to Blog
    AI & AutomationNov 20, 2026

    OpenAI GPT-Live-1.1: Voice Agents That Cost Less When Callers Pause

    Amir

    Amir Farid

    7 min read

    OpenAI GPT-Live-1.1: Voice Agents That Cost Less When Callers Pause

    OpenAI released GPT-Live-1.1 in its API on 19 November 2026. It is a full-duplex voice model, which means it can listen and speak at the same time, so callers can interrupt it mid-sentence and it answers without the pause-and-wait rhythm of older voice bots. It is the model behind voice in ChatGPT Plus and Pro since 15 October 2026, and it adds Standby mode: silent time bills at $0.02 a minute instead of $0.05. For businesses running voice agents, that is the headline: the quiet parts of a call just got 60% cheaper.

    The short version: 1.1 is a drop-in upgrade with better pronunciation, steadier accents, more accurate transcription and sharper capture of the details that break real calls — spelled names, phone numbers, order IDs and confirmation codes.

    OpenAI GPT-Live-1.1: Voice Agents That Cost Less When Callers Pause

    What GPT-Live-1.1 actually is

    GPT-Live-1.1 is OpenAI's voice model for live, spoken conversation. You connect it through the Live API at the v1/live/sessions endpoint; it handles the conversation while a backend model or agent reasons, looks up information and uses tools. Its 128,000-token context window includes audio tokens. Launch facts here come from OpenAI's developer changelog and the GPT-Live-1.1 model page.

    Full duplex, in plain English

    Full duplex means the model listens while it speaks. A caller can interrupt or correct themselves mid-sentence, and the conversation carries on naturally. Our guide to full-duplex voice AI for business explains why interruption handling is what callers notice first. New to voice agents? Start with our guide to building an AI voice agent.

    What changed from GPT-Live-1

    Support agent wearing a headset and smiling while on a call

    GPT-Live-1 reached the API on 10 September 2026. Version 1.1 improves on it; every improvement below is OpenAI's claim, not an independently verified result. OpenAI says 1.1 offers:

    • Better pronunciation and more consistent accents across languages.
    • More accurate transcription.
    • Better handling of precise details: spelled letters, phone numbers, order IDs, confirmation codes.
    • Stronger instruction following.
    • Plain-language delegation to a backend model of your choice.

    Why precise details matter on a business call

    A voice agent that mishears a phone number or an order ID creates work, not savings: the caller repeats themselves, or a human fixes it afterwards. Test detail capture first in your own pilot. Our guide to transcription for business voice covers what good capture looks like.

    Standby mode: silence at $0.02 a minute

    Standby is the new billing state — the change that moves the money. For WebRTC sessions on both GPT-Live-1 and GPT-Live-1.1, you set an idle timeout of 5 to 60 seconds. After that silence, the model pauses inference while the call stays open, resuming the moment someone speaks — at $0.02 a minute instead of $0.05. Ten idle minutes cost $0.20, against $0.50 active.

    How the 5–60 second timeout works

    You choose the timeout when you configure the session. A short timeout near 5 seconds saves more on long pauses but risks pausing while a caller is still thinking; a longer one near 60 seconds only rests during genuine waits, such as a caller looking something up. Start at 30 seconds and adjust from real recordings. Our AI calling agent cost guide shows how session pricing fits the full cost of a voice agent.

    What a call actually costs

    OpenAI's published rates are $0.05 per active minute and $0.02 per standby minute, billed per second by session duration, with no rounding up to the next minute. Backend model and tool usage is billed separately. For example:

    • A 30-minute support call, 10 minutes talking and 20 quiet: 10 × $0.05 + 20 × $0.02 = $0.90, against $1.50 if the whole call stayed active.
    • A 5-minute booking call with 2 idle minutes: 3 × $0.05 + 2 × $0.02 = $0.19.

    Take a receptionist handling 100 calls a day, averaging 4 minutes each with 1 idle minute per call. Active-only billing is 400 × $0.05 = $20.00 a day; with standby it is 300 × $0.05 + 100 × $0.02 = $17.00 — roughly $90 a month saved, before backend costs. Price the whole workflow, though — backend, telephony and infrastructure — as you would any serious AI voice calling agent budget.

    Where this changes real business calls

    Receptionists: hold the line, cheaply

    A receptionist agent spends much of a call waiting — while the caller checks a diary or reads out an address. Standby turns that waiting from $0.05 a minute into $0.02. Our AI receptionist guide for small business covers what it must handle first.

    Order-taking: get the details right

    Order calls live or die on precise details: item codes, quantities, addresses and confirmation numbers. OpenAI claims 1.1 handles spelled letters and order IDs better than 1.0 — restaurants see the same pattern in our guide to AI voice agents for restaurants.

    Support: let the caller think

    Support calls have the longest silences: the caller is reading an error message or finding an account number. Standby keeps the session alive through those gaps at the lower rate, and the 128,000-token context window means the agent still remembers the start of the call.

    Sales: never drop a warm caller

    A sales call that goes quiet is often a caller deciding, not leaving. Keeping the line open at $0.02 a minute while they think is cheap insurance against a lost lead, and delegation lets the voice agent hand a qualified caller to a backend sales agent that checks stock or pricing in real time.

    GPT-Live vs the older Realtime API

    GPT-Live is a separate path from OpenAI's older Realtime API, not a rename. Realtime agents use the /v1/realtime endpoint; GPT-Live sessions use /v1/live/sessions, with separate concurrent-session limits. The difference is the split of labour: the voice model handles the conversation while a backend model of your choice reasons and uses tools.

    The 20 January 2027 deadline

    OpenAI's published deprecation schedule removes the first-generation gpt-realtime model from the API on 20 January 2027, naming gpt-realtime-2.1 as its replacement. If your voice agent still runs on gpt-realtime, calls to it fail after that date — and moving to GPT-Live means adapting session creation and transport too, so start before the year-end change freeze. The open-source Sophon Realtime v1 voice agent and our Phonic-1 guide show the wider market moving the same way.

    How to switch without breaking live calls

    Switching models is a one-line change — set the model to gpt-live-1.1 — but switching safely is a small project. Run the new model alongside the old one on a share of live traffic, compare transcripts, and watch the two numbers that matter: how often a human steps in, and the cost per completed call. Our guide to AI agents for small business automation applies the same pilot-first discipline.

    • Week 1: route 10% of calls to GPT-Live-1.1, standby set to a 30-second timeout. Log accuracy on phone numbers, order IDs and confirmation codes against GPT-Live-1.
    • Week 2: compare cost per completed call and human-intervention rate. If 1.1 wins, move to 50%, then 100%.
    • Before you scale: set usage limits and alerts, and review delegation prompts — 1.1 uses plain-language delegation.

    Our AI agents development team builds voice agents on this stack, and our AI automation service covers the backend workflows. Compare Meta Muse voice and ElevenLabs Scribe v3, or our GPT-6.2 Astra and Sol guide. Scope a pilot with our AI agent development cost guide, or contact Coheser — our AI calling agent product is the starting point.

    Frequently asked questions

    When did GPT-Live-1.1 reach the OpenAI API?

    19 November 2026, per OpenAI's developer changelog. It had powered voice in ChatGPT Plus and Pro since 15 October; GPT-Live-1 launched in the API on 10 September 2026.

    How much does GPT-Live-1.1 cost?

    OpenAI's published rates are $0.05 per active minute and $0.02 per standby minute, billed per second; backend model and tool usage is billed separately. Ten idle minutes cost $0.20 in standby, against $0.50 active.

    What is Standby mode?

    Standby pauses inference after 5 to 60 seconds of silence while the WebRTC call stays open, resuming when someone speaks. It works on both GPT-Live-1 and GPT-Live-1.1, billing $0.02 a minute only while inference is paused.

    Should my business switch from the Realtime API now?

    If you run on gpt-realtime, plan the move now: OpenAI removes that model on 20 January 2027. Pilot GPT-Live-1.1 on a share of traffic first, and scale on the evidence.

    #GPT-Live-1.1#OpenAI#AI Voice Agents#Voice AI