
Michal Sutter
· 5 min read
OpenAI Ships GPT-Live-1.1 for Voice Agents, Cuts Idle Cost to $0.02/Min
OpenAI has made GPT-Live-1.1 available in its API under the model ID gpt-live-1.1. The GPT-Live-1.1 voice model is a full-duplex model for multilingual conversations. It can listen while it speaks. OpenAI team states that it speaks with more natural and consistent accents across languages and completes tasks more reliably. The release also adds standby, a cheaper billing state for idle sessions. For teams running voice agents, idle time on a call now costs less. You can check the deployment options and specs here through our GPULeaderboard.com chrome extension.
TL;DR
- Size: Live sessions hold a 128,000-token default context window.
- Runs on: OpenAI’s cloud only, via
v1/live/sessions. Clients connect over WebRTC or WebSocket. No open weights, no quantized variants. - Performance: OpenAI claims better accents, transcription, letter and number handling, and task completion than GPT-Live-1. No scores published.
- Best: Standby bills idle WebRTC time at $0.02 per minute versus $0.05 active, a 60% lower rate.
- Worst: OpenAI published 0 benchmark scores for 1.1. Image, video and structured outputs are not supported.
- Bottom line: Best: cheaper always-on voice sessions. Worst: quality gains are self-reported with no numbers.
What is GPT-Live-1.1?
GPT-Live-1.1 is a full-duplex speech model from OpenAI that hears and speaks in 1 stream. It follows GPT-Live-1, which reached the API on September 10, 2026. As per the OpenAI forum announcement, it has powered ChatGPT Plus and Pro voice since October 15.
Inputs and outputs are audio and text. Image and video are not supported. Streaming and function calling work. Structured outputs, fine-tuning and predicted outputs do not.
How does GPT-Live-1.1 work?
The GPT-Live family replaces the usual cascade of speech-to-text, an LLM and text-to-speech. OpenAI says the model reasons over incoming and outgoing audio together in a single model. It also handles turn detection and returns ASR transcripts natively.
Deeper work is delegated. The voice model can hand reasoning and tool calls to a backend text model. OpenAI names GPT-6 Astra for complex issues and Luna for high-volume tasks. Third-party backends also work.
The Live conversations guide sets delegation.type to client or responses. Responses mode uses OpenAI models. Client mode connects your own backend. Instructions can run up to 16,384 tokens.
What changed from GPT-Live-1?
OpenAI lists 4 improvement areas for 1.1:
- More natural and consistent accents across languages
- More accurate speech transcription
- Better handling of precise details like letters and numbers
- More reliable instruction following and task completion
How does standby mode work?
Standby pauses inference during silence while the WebRTC connection stays open. The guide documents 3 settings:
session.wake_policy.type: set tospeech, so inference waits for speechsession.idle_timeout_s: a whole number from 5 to 60 secondssession.wake_policy.vad_threshold: optional, 0 to 1, default 0.5
After the timeout, the server emits session.paused. When the user speaks, session.resumed fires. The session ID and history are kept. OpenAI warns the first reply after a short utterance may be slower. Keep the microphone track active during standby.
Standby works for both GPT-Live-1 and GPT-Live-1.1.
How much does GPT-Live-1.1 cost?
Voice sessions are billed by duration. The GPT-Live-1 model page says billing is per second. Backend model and tool usage are billed separately. That matters: a heavy backend can cost more than the voice layer.
Deployment and availability
The model is API-only. You switch an existing Live session by setting model to gpt-live-1.1. Changing the model requires a new session.
Outbound SIP calling reached all organizations on November 2, so phone agents are supported. GPT-Live-1 and GPT-Live-1.1 have separate concurrent-session limits. Each limit applies across inference regions. The 1.1 page does not publish tier numbers.
How does GPT-Live-1.1 compare?
Simulate a call</button><button class="btn gh" id="gl11rst">Reset</button></div> <pre class="log" id="gl11log">Press Simulate to walk through the standby events.</pre> <p class="note">WebRTC only. Keep the mic track active. Session ID and history are kept. First reply after a short utterance may be slower.</p> </div> <div class="pn" id="gl11p2"> <label>Session length: <span class="val" id="gl11mv">60</span> min</label> <input type="range" id="gl11min" min="1" max="480" value="60"> <label>Share of the session that is idle (in standby): <span class="val" id="gl11iv">50</span>%</label> <input type="range" id="gl11idle" min="0" max="95" value="50"> <div class="bars"> <div class="bar"><div class="lb"><span>All minutes at active rate ($0.05)</span><span class="val" id="gl11c0">$3.00</span></div><div class="tr"><div class="fl" id="gl11b0" style="background:#F0A06A"></div></div></div> <div class="bar"><div class="lb"><span>With standby ($0.02 on idle minutes)</span><span class="val" id="gl11c1">$2.10</span></div><div class="tr"><div class="fl" id="gl11b1" style="background:#7EA4FA"></div></div></div> </div> <p class="note" id="gl11sv">Voice layer only. Illustrative: the timeout window itself bills at the active rate, and backend model and tool usage are billed separately.</p> </div> <div class="pn" id="gl11p3"> <div class="grid"> <div class="card"><h4>Accents <span class="plus">+</span></h4><p>More natural and consistent accents across languages, per OpenAI.</p></div> <div class="card"><h4>Transcription <span class="plus">+</span></h4><p>Transcribes speech more accurately than GPT-Live-1, per the model page.</p></div> <div class="card"><h4>Letters and numbers <span class="plus">+</span></h4><p>Better handling of precise details like letters and numbers.</p></div> <div class="card"><h4>Task completion <span class="plus">+</span></h4><p>More reliable instruction following and task completion.</p></div> </div> <div class="bars"> <div class="bar"><div class="lb"><span>Active rate</span><span class="val">$0.05/min</span></div><div class="tr"><div class="fl gl11pb" data-w="100" style="background:#F0A06A"></div></div></div> <div class="bar"><div class="lb"><span>Standby rate (new, also on GPT-Live-1)</span><span class="val">$0.02/min</span></div><div class="tr"><div class="fl gl11pb" data-w="40" style="background:#7EA4FA"></div></div></div> </div> <p class="note">OpenAI published no benchmark scores for 1.1. All 4 gains above are self-reported.</p> </div> <div class="ft"> <span>Sources: <a href="https://developers.openai.com/api/docs/models/gpt-live-1.1" target="_blank" rel="noopener">model page</a> · <a href="https://developers.openai.com/api/docs/guides/live-conversations" target="_blank" rel="noopener">Live guide</a></span> <span class="brand">Built by Marktechpost</span> </div> </div> <script> (function(){ var R=document.getElementById('mtp-gl11'); function $(i){return document.getElementById(i)} function resize(){try{parent.postMessage({mtpFrame:'gl11',height:R.offsetHeight+40},'*')}catch(e){}} var tabs=R.querySelectorAll('.tab'); for(var i=0;i<tabs.length;i++){tabs[i].addEventListener('click',function(){ for(var j=0;j<tabs.length;j++){tabs[j].classList.remove('on');$('gl11p'+j).classList.remove('on')} this.classList.add('on');var t=this.getAttribute('data-t');$('gl11p'+t).classList.add('on'); if(t==='2')calc(); if(t==='3')priceBars(); setTimeout(resize,60); })} /* step-through */ var steps=[ {on:['gl11n0','gl11a0','gl11n1'],t:'1. Caller audio streams in. The model keeps listening even while it is speaking (full duplex).',c:'Step 1 of 3'}, {on:['gl11n1','gl11a2','gl11n2'],t:'2. For reasoning or tools, the voice model delegates to a backend model while the conversation continues.',c:'Step 2 of 3'}, {on:['gl11n2','gl11a3','gl11n1','gl11a1','gl11n0'],t:'3. The result returns and is spoken back. ASR transcripts and response text come back natively.',c:'Step 3 of 3'}]; var ids=['gl11n0','gl11n1','gl11n2','gl11a0','gl11a1','gl11a2','gl11a3'],si=-1; function showAll(){ids.forEach(function(x){$(x).classList.remove('dim')})} $('gl11next').addEventListener('click',function(){si=(si+1)%steps.length;var s=steps[si]; ids.forEach(function(x){$(x).classList.toggle('dim',s.on.indexOf(x)<0)});$('gl11st').textContent=s.t;$('gl11ct').textContent=s.c;}); $('gl11all').addEventListener('click',function(){si=-1;showAll();$('gl11st').textContent='One model handles listening and speaking. No separate speech-to-text and text-to-speech chain.';$('gl11ct').textContent='Overview'}); /* standby */ $('gl11to').addEventListener('input',function(){$('gl11tv').textContent=this.value;$('gl11s1').style.flexGrow=(1+this.value/15).toFixed(2)}); var timers=[]; function rst(){timers.forEach(clearTimeout);timers=[];for(var k=0;k<4;k++)$('gl11s'+k).classList.remove('lit');$('gl11log').innerHTML='Press Simulate to walk through the standby events.'} $('gl11rst').addEventListener('click',rst); $('gl11sim').addEventListener('click',function(){rst();var to=$('gl11to').value;var L=$('gl11log');L.innerHTML=''; var ev=[[0,'<b>session.resumed</b> first wake on speech, inference active ($0.05/min)'], [1,'silence detected, waiting '+to+' s (idle_timeout_s)'], [2,'<b>session.paused</b> standby, WebRTC call stays open ($0.02/min)'], [3,'user speaks, <b>session.resumed</b> same session ID, history restored']]; ev.forEach(function(e,n){timers.push(setTimeout(function(){$('gl11s'+e[0]).classList.add('lit');L.innerHTML+=(n?'\n':'')+e[1];resize()},n*1100))}); }); /* calculator */ function money(v){return '$'+v.toFixed(2)} function calc(){var m=+$('gl11min').value,p=+$('gl11idle').value/100;$('gl11mv').textContent=m;$('gl11iv').textContent=Math.round(p*100); var a=m*0.05,b=m*(1-p)*0.05+m*p*0.02;$('gl11c0').textContent=money(a);$('gl11c1').textContent=money(b); $('gl11b0').style.width='100%';$('gl11b1').style.width=(a?b/a*100:0)+'%'; $('gl11sv').textContent='Saves '+money(a-b)+' ('+(a?((a-b)/a*100).toFixed(0):0)+'%) on the voice layer. Illustrative: the timeout window itself bills at the active rate, and backend model and tool usage are billed separately.'} $('gl11min').addEventListener('input',calc);$('gl11idle').addEventListener('input',calc); function priceBars(){var b=R.querySelectorAll('.gl11pb');for(var k=0;k<b.length;k++){b[k].style.width='0';(function(el){setTimeout(function(){el.style.width=el.getAttribute('data-w')+'%'},50)})(b[k])}} var cards=R.querySelectorAll('.card');for(var c=0;c<cards.length;c++){cards[c].addEventListener('click',function(){this.classList.toggle('open');this.querySelector('.plus').textContent=this.classList.contains('open')?'−':'+';setTimeout(resize,420)})} window.addEventListener('load',resize);window.addEventListener('resize',resize);setTimeout(resize,300); })(); </script> </body></html>">Key Takeaways
- GPT-Live-1.1 is live in the API as
gpt-live-1.1. - Standby cuts idle WebRTC billing from $0.05 to $0.02 per minute.
- Idle timeout is configurable from 5 to 60 seconds.
- Accent, transcription and task gains are self-reported, with no scores.
- Backend reasoning and tools are billed separately from voice minutes.
Check out the model page, the Live conversations guide and the announcement. You can check the deployment options and specs here through our GPULeaderboard.com chrome extension.
Need to partner with us for promoting your GitHub Repo OR Hugging Face Page OR Product Release OR Webinar etc.? Connect with us
Original source
This story was published by MarkTechPost and written by Michal Sutter. SyncAI.news shows a preview; the complete article is on the publisher's site.
Read the full story on marktechpost.com


