Anthropic just released Claude Fable 5.2 and Claude Mythos 5.2. Claude Fable 5.2 is a frontier model from Anthropic for long-horizon coding, research and agentic work. Mythos 5.2 is the same underlying model with fewer cyber and biology safeguards. It beats Opus 5.5 on most evaluations and costs less than the 5.1 models.
TL;DR
- Size: 1M-token context window, 128K max output.
- Runs on: Hosted API only. Claude Platform, Amazon Web Services, Google Cloud, Microsoft Azure. No open weights.
- Performance: Ahead of Opus 5.5 on almost every benchmark Anthropic reported.
- Best: RoboDojo Sim 64.9%, more than double GPT-6 Astra (29.8%).
- Worst: AutomationBench 51.4%, behind Gemini 4 Carbon (53.9%).
- Bottom line: Best: Opus-level coding at roughly half the cost per task. Worst: Fable still blocks most cybersecurity work.
What is the difference between Claude Fable 5.2 and Mythos 5.2?
They are one model. Fable 5.2 adds classifiers that block high-risk cybersecurity tasks and most research biology. That lets Anthropic offer it to all users.
Mythos 5.2 drops those blocks. Access runs through the Cyber Verification Program and the Life Sciences Verification Program.
When Fable blocks a query, users get an answer from Opus 4.8 (cyber) or Opus 5 (biology). Anthropic reports about 55% fewer blocks than Fable 5.1. That splits into roughly 60% fewer in cybersecurity and 40% fewer in biology.
How does Claude Fable 5.2 perform on benchmarks?
Anthropic published one table comparing Fable 5.2 with Fable 5.1, Opus 5.5, GPT-6 Astra and Gemini 4 Carbon. All scores are self-reported. Key results:
- FrontierCode 2.0 (agentic coding): 49.1% at Xhigh effort, vs 45.5% for Opus 5.5.
- FrontierSWE v2: 69.9%, vs 65.5% for GPT-6 Astra.
- Remote Labor Index: 41.3%, vs 21.3% for Opus 5.5.
- Humanity’s Last Exam (with tools): 74.2%, vs 67.7% for Opus 5.5.
- OSWorld 2.1 (partial offline set): 89.3%.
- CVE Rediscovery (Mythos 5.2): 89.6%, vs 81.3% for Opus 5.5 without safeguards.
Fable 5.2 does not lead everywhere. Gemini 4 Carbon wins AutomationBench (53.9% vs 51.4%). Opus 5.5 holds a higher AA-Briefcase Elo (1800 vs 1785).
‘+(r[1]?r[1]+’%’:’n/a’)+’
‘;
w.appendChild(el);var f=el.querySelector(‘.fill’);setTimeout(function(){f.style.width=(r[1]?r[1]:0)+’%’;},40+j*90);});
R.querySelector(‘#mtpBnote’).textContent=B[i].note+’ Source: Anthropic launch table (self-reported).’;postH();}
// effort steps
var S=[
{t:’FrontierCode 2.0 · High effort’,b:’45.3%’,v:’vs Opus 5.5 best score 45.5%’,c:[[‘Fable High’,50],[‘Opus best’,100]],x:’About the same score for roughly half the cost per task.’},
{t:’FrontierCode 2.0 · Xhigh effort’,b:’49.1%’,v:’Highest of any released model Anthropic tested’,c:[],x:’Ahead of Opus 5.5 (45.5%) and GPT-6 Astra (42.8%) best scores.’},
{t:’WANDR · Medium effort’,b:’= Opus top’,v:’Agentic search, Perplexity benchmark’,c:[[‘Fable Med’,40],[‘Opus top’,100]],x:’Matches Opus 5.5 top score for about 40% of the cost per task.’},
{t:”Humanity’s Last Exam · Low effort”,b:’= Opus top’,v:’Expert-level reasoning’,c:[[‘Fable Low’,9],[‘Opus top’,100]],x:’Matches Opus 5.5 top score for less than one-tenth of the cost.’},
{t:’Chartography · Medium effort’,b:’66.4%’,v:’vs Opus 5.5 at Max effort 64.4%’,c:[[‘Fable Med’,9],[‘Opus Max’,100]],x:’Higher score for less than one-tenth of the cost per task.’},
{t:’BenchCAD 2.1 · Max effort’,b:’83.5′,v:’vs Opus 5.5 65.3 and Fable 5.1 44.4′,c:[],x:’Hardware engineering, scored 0 to 100 with partial credit.’}
];
var si=0,dots=R.querySelector(‘#mtpDots’);S.forEach(function(){var d=document.createElement(‘div’);d.className=”dot”;dots.appendChild(d);});
function drawStep(){var s=S[si],h=”
“+s.t+’
‘+s.b+’
‘+s.v+’
‘;
if(s.c.length){h+=’
‘+c[0]+’
‘;});h+=’
‘;}
h+=”;
var el=R.querySelector(‘#mtpStep’);el.innerHTML=h;
el.querySelectorAll(‘.cb’).forEach(function(b){setTimeout(function(){b.style.height=(b.getAttribute(‘data-h’)*0.6)+’px’;},60);});
dots.querySelectorAll(‘.dot’).forEach(function(d,k){d.classList.toggle(‘on’,k===si);});postH();}
R.querySelector(‘#mtpPrev’).onclick=function(){si=(si+S.length-1)%S.length;drawStep();};
R.querySelector(‘#mtpNext’).onclick=function(){si=(si+1)%S.length;drawStep();};
// safety
var SF=[
{d:[[‘All requests’,’55’],[‘Cybersecurity’,’60’],[‘Biology’,’40’]],max:100,u:’%’,n:’Approximate reduction in blocks vs Fable 5.1, early testing across claude.ai, Claude Code and Cowork.’,me:-1},
{d:[[‘Mythos 5.2′,’3.1’],[‘Opus 5.5′,’3.3’],[‘Sonnet 5.5′,’5.1’],[‘Mythos 5.1′,’21’],[‘Mythos 5′,’39’]],max:40,u:’%’,n:’Share of 1,800+ adversarial scenarios where the model used credentials or files placed outside its sandbox. Lower is better.’,me:0},
{d:[[‘Mythos 5.2′,’2.67’],[‘Opus 5.5′,’2.79’],[‘Sonnet 5.5′,’3.10’],[‘Mythos 5.1′,’3.28′]],max:4,u:”,n:’Automated behavioral audit, about 1,900 scenarios, scored 1 to 10. Lower is better.’,me:0}
];
var curS=0;
R.querySelectorAll(‘#mtpS .chip’).forEach(function(c){c.addEventListener(‘click’,function(){R.querySelectorAll(‘#mtpS .chip’).forEach(function(x){x.classList.remove(‘on’)});c.classList.add(‘on’);curS=+c.getAttribute(‘data-s’);drawS(curS);});});
function drawS(i){var w=R.querySelector(‘#mtpSBars’),s=SF[i];w.innerHTML=”;
s.d.forEach(function(r,j){var el=document.createElement(‘div’);el.className=”bar”;
el.innerHTML=’
‘+r[0]+’
‘+(i===0?’-‘:”)+r[1]+s.u+’
‘;
w.appendChild(el);var f=el.querySelector(‘.fill’);setTimeout(function(){f.style.width=(r[1]/s.max*100)+’%’;},40+j*90);});
R.querySelector(‘#mtpSnote’).textContent=s.n+’ Source: Anthropic launch post and system card.’;postH();}
// pricing
function fmt(v){return ‘$’+v.toLocaleString(‘en-US’,{maximumFractionDigits:2,minimumFractionDigits:2});}
function calc(){var i=+R.querySelector(‘#mtpIn’).value,o=+R.querySelector(‘#mtpOut’).value;
R.querySelector(‘#mtpInL’).textContent=i+’M’;R.querySelector(‘#mtpOutL’).textContent=o+’M’;
R.querySelector(‘#c1’).textContent=fmt(i*8+o*40);R.querySelector(‘#c2’).textContent=fmt(i*16+o*80);
R.querySelector(‘#c3’).textContent=fmt(i*4+o*20);R.querySelector(‘#c4’).textContent=fmt(i*10+o*50);}
R.querySelector(‘#mtpIn’).oninput=calc;R.querySelector(‘#mtpOut’).oninput=calc;calc();
window.addEventListener(‘load’,postH);window.addEventListener(‘resize’,postH);setTimeout(postH,300);
})();