Hours after Anthropic claims a Claude proof of the Riemann hypothesis, Stanford releases an all-human competing draft
AIvia The Decoder
Quick Read
Anthropic mathematician Levent Alpöge says he used Claude to prove the Riemann hypothesis, math's most famous open problem for 167 years, with a proof formalized in Lean that spans over 250,000 pages. Just hours later, Stanford researchers published an independent, entirely human competing proof, and mathematicians are now checking whether either one holds up. The article Hours after Anthropic claims a Claude proof of the Riemann hypothesis, Stanford releases an all-human competing draft appeared first on The Decoder.
Anthropic mathematician Levent Alpöge says he used Claude to prove the Riemann hypothesis, math’s most famous open problem for 167 years, with a proof formalized in Lean that spans over 250,000 pages. Just hours later, Stanford researchers published an independent, entirely human competing proof, and mathematicians are now checking whether either one holds up.
The article Hours after Anthropic claims a Claude proof of the Riemann hypothesis, Stanford releases an all-human competing draft appeared first on The Decoder.
OpenAI has released new prompt templates for GPT-6.2 Astra and Sol meant to fix common model behaviors such as overly brief answers, unnecessary follow-up questions, and excessive testing when coding. The API is changing as well, since editing a conversation's history can discard reasoning content and lower response quality. The article OpenAI shares new prompting tips for GPT-6.2 Astra and Sol to fix brief answers and unnecessary questions appeared first on The Decoder.
Thinking Machines Lab has released Inkling 2 Mini, an open-weight looped transformer with 26.4B total parameters. The model reuses its middle layers 3 times per token instead of adding more weights. That keeps the checkpoint near 20GB, small enough for a 32GB Mac or 1 NVIDIA DGX Spark. For developers, Inkling 2 Mini is a local option for agentic coding, tool use and document work. It ships under Apache 2.0 on Hugging Face, with hosted inference and fine-tuning on Tinker. TL;DR Size: 26.4B total, 3.4B active (5.4B effective per token with looping), up to 1M-token context. Runs on: Apple silicon with 32GB+ memory, 1 NVIDIA DGX Spark, or GPUs with 32GB+ VRAM. GGUF builds via Unsloth and llama.cpp. Performance: 20.1 on the Artificial Analysis Index, tied with Qwen3.8-27B (self-reported). Best: 79.2% on SWE-bench Verified (bash only). Worst: 15.2% on HLE-Diamond and 50.0% on MultiChallenge. Bottom line: Best: strong agentic coding from a 20GB checkpoint. Worst: weaker knowledge recall and multi-turn instruction following. What is Inkling 2 Mini? Inkling 2 Mini is a sparse Mixture-of-Experts model from Thinking Machines that loops its middle layers. It accepts text and images and returns text only. It follows Inkling 2 Small, a 594B-parameter model released on November 5. How does the looped transformer work? The model stores 24 layers, grouped into 4 blocks of 6. The first and last blocks run once. The 2 middle blocks run 3 times with shared weights. That gives 48 layer applications per token from 24 layers of weights. Each pass after the first receives a “message” from the previous pass. Thinking Machines carries it with Attention Residuals and a learned gate. Feed-forward layers use 128 experts, with 8 routed per token. Looping the middle layers gave a 1.51x to 1.91x compute-equivalent gain. Looping the last layers gave only 1.24x to 1.36x. The middle-loop variant won 27 of 28 evaluations. Looping costs training compute but saves inference. An unlooped baseline used 7.5% less training compute. It needed 140% more inference compute to match the looped model’s loss on 512k-token agentic traces. How was Inkling 2 Mini trained? Training began with 384 experts per layer, with 8 active per token. At midtraining, the team pruned two-thirds of them and kept the 128 most-used experts. A 1T-token distillation run from a larger internal Inkling model then healed the pruned model. Post-training used SFT, then RL over 2.2T tokens and 2,000 steps, plus on-policy distillation. Expert layers are trained and deployed in NVFP4. The last 4 layers use MXFP8. Thinking Machines says this removes quantization conversion loss. A research variant, TML-d24-varloop-research, trains with 1 to 5 passes. How does Inkling 2 Mini perform on benchmarks? All Inkling scores below are self-reported by Thinking Machines. Coding is the strongest area. Inkling 2 Mini scores 79.2% on SWE-bench Verified (bash only) and 77.9% on SWE-bench Pro v2 (Hard). It reaches 48.2% on DeepSWE v1.1, where Qwen’s own card lists Qwen3.8-27B at 42.2%. On Terminal Bench v2.1 it scores 71.6%, behind Qwen3.8-27B at 73.0%. Agentic search also holds up: 79.0% on BrowseComp with compaction and 82.7% on MCP Atlas. For documents and vision, it posts 92.8% on OmniDocBench v1.6 and 74.9% on MMMU Pro. On reasoning, it scores 86.6% on GPQA Diamond and 49.3% on ARC-AGI-2. Efficiency is the headline claim. On Stanford’s Intelligence per Watt metric, Thinking Machines reports 0.703 on a DGX Spark. It lists Qwen3.6-35B-A3B at 0.430 and Gemma-4-26B-A4B at 0.159. window.addEventListener("message",function(e){if(e.data&&e.data.mtpInk2mHeight){var f=document.getElementById("mtp-ink2m-frame");if(f&&e.source===f.contentWindow){f.style.height=e.data.mtpInk2mHeight+"px";}}}); How does Inkling 2 Mini compare with Qwen and Gemma? Each column uses the vendor’s own model card. Harnesses and settings differ, so treat cross-vendor gaps as directional. What hardware do you need to run Inkling 2 Mini? The checkpoint is about 20GB in Safetensors. Thinking Machines lists 3 supported setups: 1 NVIDIA DGX Spark, Apple silicon with 32GB+ memory, and other GPUs with 32GB+ VRAM. Quantized GGUF builds are available through Unsloth and llama.cpp. Serving support covers Transformers, vLLM, SGLang, Ollama, NVIDIA Dynamo and TokenSpeed. Tinker pricing for Mini is not disclosed in the launch materials. A vLLM quick start takes 1 command: vllm serve "thinkingmachines/Inkling-2-Mini". Where does Inkling 2 Mini fall short? Knowledge depends mostly on total parameters, and looping did not help knowledge benchmarks. Inkling 2 Mini scores -8.6 on AA-Omniscience, with a 63.4% non-hallucination rate. It is not the best small model for multi-turn chat or constraint-heavy prompts. It scores 50.0% on MultiChallenge and 33.4% on ComplexConstraints v1.1. It also trails larger models on HLE-Diamond (15.2%) and CritPt v2. Key Takeaways Inkling 2 Mini loops 24 stored layers into 48 layer passes per token. 26.4B total, 3.4B active, ~20GB, Apache 2.0, up to 1M context. Runs on a 32GB Mac or 1 DGX Spark via Ollama, vLLM or llama.cpp. Scores 79.2% on SWE-bench Verified and ties Qwen3.8-27B at 20.1 AA Index. Weak spots: knowledge recall, multi-turn chat and constraint-heavy prompts. Check out the technical blog, model card, model weights and Tinker Playground. You can check the deployment options and specs here through our GPULeaderboard.com chrome extension. Also, feel free to follow us on Twitter and don’t forget to join our 150k+ML SubReddit and Subscribe to our Newsletter. Wait! are you on telegram? now you can join us on telegram as well. Need to partner with us for promoting your GitHub Repo OR Hugging Face Page OR Product Release OR Webinar etc.? Connect with us The post Meet Thinking Machines Inkling 2 Mini: 26B Looped MoE Runs on a 32GB Mac appeared first on MarkTechPost.
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 to speech, so inference waits for speech session.idle_timeout_s: a whole number from 5 to 60 seconds session.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? <iframe id="mtp-gl11-frame" title="GPT-Live-1.1 interactive explainer" loading="lazy" style="width:100%;height:600px;border:0;display:block;background:transparent;" srcdoc="<title>GPT-Live-1.1 explained | Marktechpost</title>html,body{margin:0;padding:0;background:transparent} #mtp-gl11{max-width:880px;margin:0 auto;background:#0d0d0d!important;color:#f2f2f2!important;border:1px solid #2a2a2a!important;border-radius:16px;font-family:Inter,system-ui,-apple-system,Segoe UI,Roboto,sans-serif;overflow:hidden;box-sizing:border-box} #mtp-gl11 *{box-sizing:border-box} #mtp-gl11 hr,#mtp-gl11 p:empty,#mtp-gl11 del,#mtp-gl11 s{display:none!important} #mtp-gl11 .hd{padding:22px 24px 14px;background:radial-gradient(circle at 92% 0%,rgba(240,160,106,.25),transparent 45%),radial-gradient(circle at 70% 120%,rgba(91,140,247,.22),transparent 50%)!important} #mtp-gl11 .tag{display:inline-block;font-size:11px;letter-spacing:.08em;text-transform:uppercase;color:#0d0d0d!important;background:#f2f2f2!important;border-radius:99px;padding:3px 10px;font-weight:700} #mtp-gl11 h3{margin:10px 0 4px;font-size:22px;line-height:1.25;color:#fff!important} #mtp-gl11 .sub{margin:0;color:#a9a9a9!important;font-size:14px} #mtp-gl11 .tabs{display:flex;gap:6px;padding:0 24px;border-bottom:1px solid #262626!important;flex-wrap:wrap} #mtp-gl11 .tab{background:transparent!important;color:#9a9a9a!important;border:0!important;border-bottom:2px solid transparent!important;padding:12px 10px;font:600 14px Inter,system-ui,sans-serif;cursor:pointer} #mtp-gl11 .tab.on{color:#fff!important;border-bottom-color:#F0A06A!important} #mtp-gl11 .pn{display:none;padding:20px 24px 8px;animation:gl11fade .35s ease} #mtp-gl11 .pn.on{display:block} @keyframes gl11fade{from{opacity:0;transform:translateY(6px)}to{opacity:1;transform:none}} #mtp-gl11 .row{display:flex;gap:10px;align-items:center;flex-wrap:wrap} #mtp-gl11 .btn{background:#f2f2f2!important;color:#0d0d0d!important;border:0!important;border-radius:99px;padding:8px 16px;font:700 13px Inter,system-ui,sans-serif;cursor:pointer} #mtp-gl11 .btn.gh{background:transparent!important;color:#f2f2f2!important;border:1px solid #3a3a3a!important} #mtp-gl11 .note{font-size:12px;color:#8a8a8a!important;margin:10px 0} #mtp-gl11 .box{background:#151515!important;border:1px solid #2a2a2a!important;border-radius:12px;padding:14px} #mtp-gl11 svg text{font-family:Inter,system-ui,sans-serif} #mtp-gl11 .node{transition:all .4s} #mtp-gl11 .flow{stroke-dasharray:6 6;animation:gl11dash 1s linear infinite} @keyframes gl11dash{to{stroke-dashoffset:-12}} #mtp-gl11 .pulse{animation:gl11pulse 1.6s ease-in-out infinite;transform-origin:center;transform-box:fill-box} @keyframes gl11pulse{0%,100%{transform:scale(1);opacity:1}50%{transform:scale(1.08);opacity:.85}} #mtp-gl11 .dim{opacity:.25} #mtp-gl11 .stepTxt{min-height:44px;font-size:14px;color:#ddd!important;margin:8px 0} #mtp-gl11 .tl{display:flex;height:34px;border-radius:8px;overflow:hidden;border:1px solid #2a2a2a!important;margin:12px 0 6px} #mtp-gl11 .seg{display:flex;align-items:center;justify-content:center;font-size:11px;font-weight:700;color:#0d0d0d!important;transition:flex-grow .6s,opacity .4s;opacity:.2;white-space:nowrap;overflow:hidden} #mtp-gl11 .seg.lit{opacity:1} #mtp-gl11 .log{font:12px/1.6 ui-monospace,Menlo,Consolas,monospace;background:#0a0a0a!important;color:#cfcfcf!important;border:1px solid #232323!important;border-radius:8px;padding:10px;min-height:96px;margin:0;white-space:pre-wrap;word-break:break-word} #mtp-gl11 .log b{color:#F0A06A!important} #mtp-gl11 input[type=range]{width:100%;accent-color:#F0A06A} #mtp-gl11 label{font-size:13px;color:#bbb!important;display:block;margin:10px 0 4px} #mtp-gl11 .val{color:#fff!important;font-weight:700} #mtp-gl11 .bars{margin-top:14px} #mtp-gl11 .bar{margin:8px 0} #mtp-gl11 .bar .lb{display:flex;justify-content:space-between;font-size:13px;margin-bottom:4px;color:#ddd!important} #mtp-gl11 .bar .tr{height:16px;background:#1c1c1c!important;border-radius:99px;overflow:hidden} #mtp-gl11 .bar .fl{height:100%;width:0;border-radius:99px;transition:width .8s cubic-bezier(.2,.8,.2,1)} #mtp-gl11 .grid{display:grid;grid-template-columns:1fr 1fr;gap:10px;align-items:start} #mtp-gl11 .card{background:#151515!important;border:1px solid #2a2a2a!important;border-radius:12px;padding:12px 14px;cursor:pointer;transition:border-color .25s,transform .25s} #mtp-gl11 .card:hover{border-color:#5B8CF7!important;transform:translateY(-2px)} #mtp-gl11 .card h4{margin:0;font-size:14px;color:#fff!important} #mtp-gl11 .card p{margin:6px 0 0;font-size:13px;color:#aaa!important;max-height:0;overflow:hidden;transition:max-height .4s} #mtp-gl11 .card.open p{max-height:120px} #mtp-gl11 .card .plus{float:right;color:#F0A06A!important;font-weight:700} #mtp-gl11 code{background:#1f1f1f!important;color:#F0A06A!important;border:0!important;padding:1px 5px;border-radius:4px;font-size:12px} #mtp-gl11 .ft{display:flex;justify-content:space-between;gap:10px;flex-wrap:wrap;padding:12px 24px 16px;border-top:1px solid #222!important;font-size:12px;color:#777!important;margin-top:12px} #mtp-gl11 .ft a{color:#9a9a9a!important} #mtp-gl11 .ft .brand{color:#76B900!important;font-weight:700} @media (max-width:640px){#mtp-gl11 .hd{padding:18px 16px 12px}#mtp-gl11 h3{font-size:19px}#mtp-gl11 .tabs{padding:0 12px}#mtp-gl11 .tab{font-size:12px;padding:10px 6px}#mtp-gl11 .pn{padding:16px}#mtp-gl11 .grid{grid-template-columns:1fr}#mtp-gl11 .seg{font-size:9px}#mtp-gl11 .ft{padding:12px 16px}} <div id="mtp-gl11"> <div class="hd"> <span class="tag">OpenAI · gpt-live-1.1</span> <h3>GPT-Live-1.1, explained</h3> <p class="sub">Full-duplex voice model in the API. $0.05/min active, $0.02/min standby.</p> </div> <div class="tabs" role="tablist"> <button class="tab on" data-t="0">How it works</button> <button class="tab" data-t="1">Standby mode</button> <button class="tab" data-t="2">Cost calculator</button> <button class="tab" data-t="3">What changed</button> </div> <div class="pn on" id="gl11p0"> User audio WebRTC / WebSocket / SIP GPT-Live-1.1 listens while speaking Backend model Astra · Luna · 3rd party speech in speech out delegate tools result <div class="stepTxt" id="gl11st">One model handles listening and speaking. No separate speech-to-text and text-to-speech chain.</div> <div class="row"> <button class="btn" id="gl11next">Step through ▸</button> <button class="btn gh" id="gl11all">Show all</button> <span class="note" id="gl11ct" style="margin:0">Overview</span> </div> <p class="note">Delegation modes: <code>delegation.type</code> = <code>client</code> (your backend) or <code>responses</code> (OpenAI models). Default context: 128,000 tokens.</p> </div> <div class="pn" id="gl11p1"> <label>Idle timeout <code>session.idle_timeout_s</code>: <span class="val" id="gl11tv">20</span> s (allowed 5 to 60)</label> <div class="tl"> <div class="seg" id="gl11s0" style="flex-grow:3;background:#F0A06A">Active · talking</div> <div class="seg" id="gl11s1" style="flex-grow:2;background:#d9c2b0">Silence (timeout)</div> <div class="seg" id="gl11s2" style="flex-grow:4;background:#7EA4FA">Standby · $0.02/min</div> <div class="seg" id="gl11s3" style="flex-grow:2;background:#F0A06A">Resumed</div> </div> <div class="row" style="margin:8px 0 10px"><button class="btn" id="gl11sim"> 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> <label>Share of the session that is idle (in standby): <span class="val" id="gl11iv">50</span>%</label> <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=" target=" rel="noopener">model page</a> · <a href=" target=" rel="noopener">Live guide</a></span> <span class="brand">Built by Marktechpost</span> </div> </div> (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); })(); "> window.addEventListener("message",function(e){if(e.data&&e.data.mtpFrame==="gl11"&&e.data.height){document.getElementById("mtp-gl11-frame").style.height=e.data.height+"px";}}); 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. Also, feel free to follow us on Twitter and don’t forget to join our 150k+ML SubReddit and Subscribe to our Newsletter. Wait! are you on telegram? now you can join us on telegram as well. Need to partner with us for promoting your GitHub Repo OR Hugging Face Page OR Product Release OR Webinar etc.? Connect with us The post OpenAI Ships GPT-Live-1.1 for Voice Agents, Cuts Idle Cost to $0.02/Min appeared first on MarkTechPost.