Google Releases Gemini 4 Argon API With 1M Token Output Restrict

Google DeepMind has made Gemini 4 Argon usually out there to builders, enterprises and paid Gemini subscribers. Gemini 4 Argon is a sparse mixture-of-experts frontier mannequin from Google constructed for lengthy, multi-step work. Its goal domains are coding, finance, authorized, science and cybersecurity. Google first introduced Argon on September 30 for a small group of cyber defenders. It now reaches the paid Gemini API at $2 enter and $10 output per million tokens. The important thing change for builders is output size: Argon can write as much as 1M tokens in a single response.

TL;DR

Measurement: Parameter depend not disclosed (sparse MoE). 1M token window shared by enter and output.
Runs on: Google’s cloud solely, through Gemini API, AI Studio and Gemini platform. Skilled on TPUs. No open weights.
Efficiency: Leads Google’s desk on data work, lengthy context and video. Combined on agentic coding.
Greatest: 91.7% on LVBench (lengthy video understanding).
Worst: 55.0% on FrontierSWE v2, final of 4 fashions in Google’s desk.

What’s Gemini 4 Argon?

In response to the mannequin card, Argon is a brand new mannequin, not a fine-tune of an earlier Gemini. It’s a large-scale sparse MoE transformer with native multimodal help. It accepts textual content, photos, audio and video, and it outputs textual content. Google skilled it on TPUs utilizing JAX and ML Pathways. The data cutoff is March 2026. Google has not disclosed whole or lively parameter counts.

How does the 1M token output restrict work?

Google raised the output restrict to 1M tokens, up from 64K on earlier Gemini fashions. The mannequin card describes a mixed enter plus output window of as much as 1M tokens. So an extended immediate reduces the room left for output. Google paperwork a brand new continuation protocol for lengthy decodes in its developer docs and enterprise docs. At introductory pricing, a full 1M token response prices $10 in output tokens.

How does Gemini 4 Argon carry out on benchmarks?

Google compares Argon with GPT-6 Astra, Claude Fable 5.1 and Claude Opus 5.5. Argon posts the highest rating on most rows.

DeepSWE v1.1: 77.9%, forward of Opus 5.5 (74.2%) and GPT-6 Astra (74.1%).
Vals Index: 68.90%, ranked #1 by Vals AI forward of Claude Sonnet 5.5 (67.04%).
AutomationBench: 51.3%, versus 42.5% for Opus 5.5.
Harvey’s Authorized Agent Benchmark: 19.6%, practically 3x the subsequent mannequin, however nonetheless a low absolute rating.
GraphWalks, 256K to 1M tokens: 84.2%, versus 71.8% for GPT-6 Astra.
CWE-bench v1: 68.0%, tied with GPT-6 Astra.

Google additionally stories 87% on the ERQA embodied reasoning benchmark. Learn the analysis methodology earlier than drawing conclusions. Most rival scores are self-reported or pulled from public leaderboards. Argon’s DeepSWE rating is self-computed with a mini-swe-agent harness. The rival DeepSWE numbers come from system playing cards and a leaderboard.

The place does Gemini 4 Argon fall brief?

Argon trails on 5 rows of Google’s personal desk:

FrontierSWE v2: 55.0%, versus 65.5% for GPT-6 Astra.
Terminal-Bench 4.0: 57.4%, versus 66.4% for Opus 5.5.
Terminal-Bench Science 0.1: 57.6%, versus 68.1% for GPT-6 Astra.
PostTrainBench: 45.3%, versus 49.3% for Opus 5.5.
OSWorld-2.0 (offline subset): 69.2%, versus 72.6% for GPT-6 Astra.

The mannequin card additionally lists hallucinations and occasional slowness or timeouts. Beneath Google’s Frontier Security Framework, Google says it can not rule out the organic and cybersecurity crucial functionality ranges. It stories deploying safeguards it assesses as reducing that danger.

<div class="pane on" id="p-how">
<svg viewBox="0 0 820 170" width="100%" type="show:block;margin-bottom:12px" aria-label="Sparse combination of consultants routing diagram">
<defs><linearGradient id="gg" x1="0" x2="1"><cease offset="0" stop-color="#4285F4"/><cease offset="1" stop-color="#9168F0"/></linearGradient></defs>
<g font-size="12" fill="#9AA0A6">
<rect x="10" y="20" width="120" peak="26" rx="6" fill="#161B24" stroke="#273041"/><textual content x="70" y="37" text-anchor="center" fill="#E8EAED">Textual content</textual content>
<rect x="10" y="52" width="120" peak="26" rx="6" fill="#161B24" stroke="#273041"/><textual content x="70" y="69" text-anchor="center" fill="#E8EAED">Photographs</textual content>
<rect x="10" y="84" width="120" peak="26" rx="6" fill="#161B24" stroke="#273041"/><textual content x="70" y="101" text-anchor="center" fill="#E8EAED">Audio</textual content>
<rect x="10" y="116" width="120" peak="26" rx="6" fill="#161B24" stroke="#273041"/><textual content x="70" y="133" text-anchor="center" fill="#E8EAED">Video</textual content>
</g>
<path d="M130 81 H230" stroke="url(#gg)" stroke-width="2" stroke-dasharray="6 6"><animate attributeName="stroke-dashoffset" from="24" to="0" dur="1s" repeatCount="indefinite"/></path>
<rect x="230" y="55" width="110" peak="52" rx="10" fill="#1A2233" stroke="#4285F4"/>
<textual content x="285" y="78" text-anchor="center" font-size="13" font-weight="700" fill="#fff">Router</textual content>
<textual content x="285" y="95" text-anchor="center" font-size="11" fill="#9AA0A6">per token</textual content>
<g id="mtp-exp"></g>
<rect x="640" y="55" width="170" peak="52" rx="10" fill="#1A2233" stroke="#34A853"/>
<textual content x="725" y="78" text-anchor="center" font-size="13" font-weight="700" fill="#fff">Textual content output</textual content>
<textual content x="725" y="95" text-anchor="center" font-size="11" fill="#9AA0A6">as much as 1M tokens</textual content>
</svg>
<div class="circulate">
<div class="st on" data-s="0"><div class="n">STEP 1</div><div class="t">Multimodal enter</div></div>
<div class="st" data-s="1"><div class="n">STEP 2</div><div class="t">Sparse MoE routing</div></div>
<div class="st" data-s="2"><div class="n">STEP 3</div><div class="t">Lengthy reasoning</div></div>
<div class="st" data-s="3"><div class="n">STEP 4</div><div class="t">1M token output</div></div>
</div>
<div class="det" id="mtp-det"></div>
<div class="ctrl"><button class="btn" id="mtp-play"> Step via</button><button class="btn g" id="mtp-next">Subsequent step</button></div>
<div class="win">
<div type="font-size:13px;font-weight:700;margin-bottom:8px">Shared 1M token window: drag to set your immediate dimension</div>
<enter sort="vary" id="mtp-in" min="0" max="1000" worth="200" aria-label="Enter tokens in hundreds">
<div class="bar"><i id="mtp-bi" type="background:#4285F4"></i><i id="mtp-bo" type="background:#34A853"></i></div>
<div class="lab"><span id="mtp-li"></span><span id="mtp-lo"></span></div>
</div>
</div>

<div class="pane" id="p-price">
<div class="seg"><button class="on" data-p="intro">Intro (to Nov 30, 2026)</button><button data-p="std">Customary (after)</button></div>
<div class="calc">
<div><label>Enter tokens per 30 days: <output id="mtp-ci"></output></label><enter sort="vary" id="mtp-rin" min="1" max="500" worth="100"></div>
<div><label>Output tokens per 30 days: <output id="mtp-co"></output></label><enter sort="vary" id="mtp-rout" min="1" max="200" worth="20"></div>
</div>
<div id="mtp-prows"></div>
<div class="be aware">Record costs per 1M tokens: Argon $2/$10 intro, $4/$20 after Nov 30, 2026. GPT-6 Astra $10/$50 (prompts below 272K). Claude Opus 5.5 $4/$20. Claude Fable 5.1 $10/$50. Estimate excludes caching, batch and long-context surcharges.</div>
</div>

<div class="pane" id="p-access">
<div class="grid">
<div class="acc"><h4><span class="dot" type="background:#4285F4"></span>Builders</h4><p>Gemini API paid tier in Google AI Studio and Android Studio. Mannequin ID: <code>gemini-4-argon</code>.</p></div>
<div class="acc"><h4><span class="dot" type="background:#9168F0"></span>Enterprises</h4><p>Gemini platform, Gemini for enterprise and Google Antigravity.</p></div>
<div class="acc"><h4><span class="dot" type="background:#FBBC04"></span>Customers</h4><p>Gemini app and AI Mode in Seek for Google AI Professional and Extremely subscribers.</p></div>
<div class="acc"><h4><span class="dot" type="background:#EA4335"></span>Cyber defenders</h4><p>Fairwind Program companions get Argon with out cyber guardrails, on a rolling foundation.</p></div>
</div>
<div class="be aware">Identified limits from the mannequin card: hallucinations, occasional slowness or timeouts, data cutoff March 2026. Parameter depend not disclosed.</div>
</div>

<div class="ft"><span>Sources: Google weblog, Gemini 4 Argon mannequin card and eval methodology, Vals AI. Checked Oct 13, 2026.</span><span class="mtp">© Marktechpost</span></div>
</div>
<script>
(perform(){
var R=doc.getElementById(‘mtp-g4a’);
perform $(s){return R.querySelector(s)} perform $$(s){return R.querySelectorAll(s)}
perform resize(){strive{father or mother.postMessage({mtpH:R.offsetHeight+40,id:’mtp-g4a’},’*’)}catch(e){}}
// tabs
$$(‘.tab’).forEach(perform(b){b.onclick=perform(){$$(‘.tab’).forEach(perform(x){x.classList.take away(‘on’)});$$(‘.pane’).forEach(perform(x){x.classList.take away(‘on’)});b.classList.add(‘on’);$(‘#p-‘+b.dataset.t).classList.add(‘on’);if(b.dataset.t===’bench’)drawB(cur);if(b.dataset.t===’worth’)drawP();setTimeout(resize,60)}});
// consultants svg
var g=$(‘#mtp-exp’),NS=’http://www.w3.org/2000/svg’,ex=[];
for(var i=0;i<8;i++){var y=12+i*19;var l=doc.createElementNS(NS,’path’);l.setAttribute(‘d’,’M340 81 C 380 81 380 ‘+(y+7)+’ 420 ‘+(y+7));l.setAttribute(‘stroke’,’#273041′);l.setAttribute(‘fill’,’none’);l.setAttribute(‘class’,’exp’);g.appendChild(l);
var r=doc.createElementNS(NS,’rect’);r.setAttribute(‘x’,420);r.setAttribute(‘y’,y);r.setAttribute(‘width’,150);r.setAttribute(‘peak’,14);r.setAttribute(‘rx’,4);r.setAttribute(‘fill’,’#202734′);r.setAttribute(‘class’,’exp’);g.appendChild(r);
var o=doc.createElementNS(NS,’path’);o.setAttribute(‘d’,’M570 ‘+(y+7)+’ C 605 ‘+(y+7)+’ 605 81 640 81′);o.setAttribute(‘stroke’,’#273041′);o.setAttribute(‘fill’,’none’);o.setAttribute(‘class’,’exp’);g.appendChild(o);ex.push([l,r,o]);}
var t=doc.createElementNS(NS,’textual content’);t.setAttribute(‘x’,495);t.setAttribute(‘y’,10);t.setAttribute(‘text-anchor’,’center’);t.setAttribute(‘font-size’,’11’);t.setAttribute(‘fill’,’#9AA0A6′);t.textContent=”Consultants (just a few hearth per token)”;g.appendChild(t);
var cols=[‘#4285F4′,’#9168F0’];
perform route(){var a=Math.flooring(Math.random()*8),b=(a+1+Math.flooring(Math.random()*7))%8;ex.forEach(perform(e,ok)ok===b);e[0].setAttribute(‘stroke’,on?cols[k===a?0:1]:’#273041′);e[2].setAttribute(‘stroke’,on?cols[k===a?0:1]:’#273041′);e[1].setAttribute(‘fill’,on?cols[k===a?0:1]:’#202734′))}
route();setInterval(route,900);
// steps
var D=[‘<b>Natively multimodal input.</b> Argon takes text, images, audio and video in one request. Google says it is a new model, not a fine-tune of an earlier Gemini.’,
‘<b>Sparse mixture of experts.</b> A router sends each token to a small subset of expert networks, so only part of the model runs per token. Google has not disclosed total or active parameters.’,
‘<b>Built for long-horizon work.</b> Google trained Argon with reinforcement learning on multi-step reasoning, problem-solving and theorem-proving data. It targets coding, finance, legal, science and cyber defense.’,
‘<b>Output up to 1M tokens.</b> The output limit rose from 64K to 1M tokens. Input and output share one 1M window, and long responses use a new continuation protocol.’];
var s=0,timer=null;perform step(n){s=n;$$(‘.st’).forEach(perform(x){x.classList.toggle(‘on’,+x.dataset.s===n)});$(‘#mtp-det’).innerHTML=D[n];resize()}
$$(‘.st’).forEach(perform(x){x.onclick=perform(){cease();step(+x.dataset.s)}});
perform cease(){if(timer){clearInterval(timer);timer=null;$(‘#mtp-play’).textContent=” Step via”}}
$(‘#mtp-next’).onclick=perform(){cease();step((s+1)%4)};
$(‘#mtp-play’).onclick=perform(){if(timer){cease();return}$(‘#mtp-play’).textContent=”❚❚ Pause”;timer=setInterval(perform(){step((s+1)%4)},2600)};
step(0);
// window
perform win(){var v=+$(‘#mtp-in’).worth;$(‘#mtp-bi’).type.width=(v/10)+’%’;$(‘#mtp-bo’).type.width=((1000-v)/10)+’%’;$(‘#mtp-li’).textContent=”Enter: “+v+’Okay tokens’;$(‘#mtp-lo’).textContent=”Room left for output: “+(1000-v)+’Okay tokens’}
$(‘#mtp-in’).oninput=win;win();
// bench
var M=[‘Gemini 4 Argon’,’GPT-6 Astra’,’Claude Opus 5.5′,’Claude Fable 5.1′],C=[‘#4285F4′,’#9AA0A6′,’#C58AF9′,’#F28B82′];
var B=[
{n:’Vals Index’,d:2,v:[68.90,63.13,66.97,65.83],src:’Impartial: Vals AI leaderboard (68.90%, #1).’},
{n:’DeepSWE v1.1′,v:[77.9,74.1,74.2,67.4],src:’Argon self-computed with a mini-swe-agent harness; rivals from system playing cards and leaderboard.’},
{n:’AutomationBench’,v:[51.3,41.4,42.5,31.4],src:’Zapier public leaderboard, personal set.’},
{n:’LVBench’,v:[91.7,87.5,83.7,79.7],src:’Self-computed with out instruments. Gemini sampled at 1 FPS.’},
{n:’GraphWalks 256K to 1M’,v:[84.2,71.8,66.8,65.0],src:’Self-computed for all fashions on 200 issues.’},
{n:’CWE-bench v1′,v:[68.0,68.0,67.0,58.0],src:’Official public leaderboard. Argon ties GPT-6 Astra.’},
{n:’FrontierSWE v2′,v:[55.0,65.5,62.3,56.3],src:’Proximal public leaderboard. Argon is final of 4.’,lose:1},
{n:’Terminal-Bench 4.0′,v:[57.4,58.2,66.4,57.9],src:’Argon self-computed; rivals from the general public leaderboard. Argon is final of 4.’,lose:1}];
var cur=0,ch=$(‘#mtp-chips’);
B.forEach(perform(b,i){var c=doc.createElement(‘button’);c.className=”chip”+(i===0?’ on’:”)+(b.lose?’ lose’:”);c.textContent=b.n;c.onclick=perform(){$$(‘.chip’).forEach(perform(x){x.classList.take away(‘on’)});c.classList.add(‘on’);drawB(i)};ch.appendChild(c)});
var rows=$(‘#mtp-rows’);M.forEach(perform(m,i){rows.insertAdjacentHTML(‘beforeend’,'<div class="row"><span class="nm">’+m+'</span><div class="observe"><div class="fill" type="background:’+C[i]+’"></div></div><span class="val">0%</span></div>’)});
perform drawB(i){cur=i;var b=B[i],f=$$(‘#mtp-rows .fill’),v=$$(‘#mtp-rows .val’);f.forEach(perform(x){x.type.width=”0″});setTimeout(perform(){b.v.forEach(perform(x,ok))},30);$(‘#mtp-bnote’).textContent=b.src+’ Dashed chips mark benchmarks the place Argon trails. Scores from Google’s eval desk until famous.’}
// worth
var P={intro:[[2,10],[10,50],[4,20],[10,50]],std:[[4,20],[10,50],[4,20],[10,50]]},pm=’intro’;
$$(‘.seg button’).forEach(perform(b){b.onclick=perform(){$$(‘.seg button’).forEach(perform(x){x.classList.take away(‘on’)});b.classList.add(‘on’);pm=b.dataset.p;drawP()}});
var pr=$(‘#mtp-prows’);M.forEach(perform(m,i){pr.insertAdjacentHTML(‘beforeend’,'<div class="row"><span class="nm">’+m+'</span><div class="observe"><div class="fill" type="background:’+C[i]+’"></div></div><span class="val">$0</span></div>’)});
perform drawP(){var a=+$(‘#mtp-rin’).worth,o=+$(‘#mtp-rout’).worth;$(‘#mtp-ci’).textContent=a+’M’;$(‘#mtp-co’).textContent=o+’M’;var c=P[pm].map(perform(p){return a*p[0]+o*p[1]}),mx=Math.max.apply(null,c),f=$$(‘#mtp-prows .fill’),v=$$(‘#mtp-prows .val’);c.forEach(perform(x,ok){f[k].type.width=(x/mx*100)+’%’;v[k].textContent=”$”+Math.spherical(x).toLocaleString()})}
$(‘#mtp-rin’).oninput=drawP;$(‘#mtp-rout’).oninput=drawP;
window.addEventListener(‘load’,resize);window.addEventListener(‘resize’,resize);setTimeout(resize,300);
})();
</script>
</physique></html>
” type=”width:100%;peak:600px;border:0;show:block;background:clear;” loading=”lazy” scrolling=”no”>

(perform(){var f=doc.getElementById(‘mtp-g4a-frame’);window.addEventListener(‘message’,perform(e){if(e.knowledge&&e.knowledge.id===’mtp-g4a’&&e.knowledge.mtpH&&e.supply===f.contentWindow){f.type.peak=e.knowledge.mtpH+’px’;}});})();

Gemini 4 Argon vs GPT-6 Astra vs Claude Opus 5.5 vs Claude Fable 5.1

Function	Gemini 4 Argon	GPT-6 Astra	Claude Opus 5.5	Claude Fable 5.1
Structure	Sparse MoE transformer	Not disclosed	Not disclosed	Not disclosed
Whole / lively params	Not disclosed	Not disclosed	Not disclosed	Not disclosed
Context window	1M (enter + output mixed)	1,050,000	1M	1M
Max output	As much as 1M	128K	128K	128K
Enter modalities	Textual content, picture, audio, video	Textual content, picture	Textual content, picture	Textual content, picture
Information cutoff	Mar 2026	Apr 30, 2026	Jun 2026	Jun 2026
API worth (in / out per 1M)	$2 / $10 intro, $4 / $20 after Nov 30	$10 / $50	$4 / $20	$10 / $50
License / entry	Proprietary, API solely	Proprietary, API solely	Proprietary, API solely	Proprietary, API solely
Vals Index	68.90%	63.13%	66.97%	65.83%
DeepSWE v1.1 (Google desk)	77.9%	74.1%	74.2%	67.4%
Terminal-Bench 4.0 (Google desk)	57.4%	58.2%	66.4%	57.9%
FrontierSWE v2 (Google desk)	55.0%	65.5%	62.3%	56.3%

What did Google use Argon for earlier than launch?

Google shared inner ends in its launch submit:

Argon in Antigravity raised agentic code submissions by over 35% and minimize rollbacks by practically a 3rd.
It delivered a 2.85x end-to-end vLLM inference speedup for Gemma 4 E2B on an A100.
In Google’s open-source Agent Growth Equipment, it minimize session prep time by 40%. Chilly imports received 21% sooner.
On Grey Swan’s oblique immediate injection benchmark, assault success fell to 0.7% throughout 15 adaptive makes an attempt.

How a lot does Gemini 4 Argon price, and the place can you employ it?

Introductory pricing is $2 enter and $10 output per million tokens. Cached enter tokens are 95% off. After November 30, 2026, the value rises to $4 and $20. That matches Claude Opus 5.5 and sits beneath GPT-6 Astra’s $10 and $50.

Entry paths, per the developer submit:

Builders: Gemini API paid tier in AI Studio and Android Studio.
Enterprises: Gemini platform, Gemini for enterprise and Google Antigravity.
Customers: Gemini app and AI Mode in Seek for Google AI Professional and Extremely subscribers.
Cyber defenders: the Fairwind Program, with entry to Argon with out cyber guardrails.