Anthropic Releases Claude Opus 5.5: Fable 5.1-Level Performance at 40% Lower Running Cost Than Opus 5

Anthropic has released Claude Opus 5.5, the first model in its new Claude 5.5 family. The team states it performs at the level of Claude Fable 5.1 on most work. It also costs 40% less to run than Opus 5 on typical workloads at default settings. On Anthropic’s own benchmarks, it leads in agentic coding, computer use, and knowledge work.

Is it deployable? Yes, as a managed API model. Anthropic has not released weights, so self-hosting is not an option. Developers can call claude-opus-5-5 on the Claude Platform, Amazon Web Services, Google Cloud, and Microsoft Azure. Zero data retention is available, as with previous Opus models.

Benchmarks: Strong Lead, Not a Clean Sweep

Opus 5.5 scores use adaptive thinking at max effort, with production safeguards enabled.

Terminal-Bench 4.0 is reported at xhigh effort for Opus 5.5. GPT-6 Astra still leads on Terminal-Bench-Science and AutomationBench. Zapier ran AutomationBench without fallback models, so safeguard interventions counted as failures. Anthropic also cautions that benchmark margins are becoming a less reliable guide. In its own use, the gap to Fable 5.1 is narrower than the scores suggest.

The cost-adjusted results are more telling. At default (medium) effort, Opus 5.5 scores 54.6% on FrontierCode. That beats GPT-6 Astra’s top score of 53.3% at about a fifth of the cost per task. On CursorBench, medium effort scores 52.5%. That is 11 points above GPT-5.6 Sol’s best, at about a third of the cost.

Pricing and Speed

Opus 5.5 needs less compute to serve than Opus 5, and pricing reflects that.

Cache reads make up most agentic and coding costs, and they drop 60%. Opus 5.5 also uses fewer tokens per task. Together, that nets out to the 40% cost reduction. Output generation is more than 30% faster than Opus 5. Fast mode in Claude Code and the Claude Platform offers up to 2.5x speed at $8 input and $40 output per million tokens.

Anthropic is also raising five-hour usage limits on Pro, Max, Team, and seat-based Enterprise plans. Subscribers get a rate limit reset they can save and use later.

What Early Testers Reported

• One tester completed a 680,000-line code migration in less than a day.
• Another audited and fixed a 200,000-line codebase in under 3 hours. Opus 5 took over 20 hours and 2.5x the tokens.
• In an internal C to Rust port of HAProxy, Opus 5.5 finished in 9.5 hours. Fable 5.1 took 12 hours, and Opus 5.5 cost 51% less.
• Deloitte says Opus 5.5 at lowest effort caught 72% of known review bugs. Opus 5 at high effort caught 56%.
• In a hard-to-source earnings report test, 16 of 18 Opus 5.5 reports cleared Anthropic’s quality bar. Fable 5.1 and Opus 5 never did.

Writing style also changed. Opus 5.5 puts key information first, uses less jargon, and follows the writing rules you give it.

Safety, Safeguards, and API Changes

Opus 5.5 is Anthropic’s first release since CEO Dario Amodei called for pacing the frontier. External evaluators including METR and Frontier Design tested it before release. It posts the best score to date on Anthropic’s automated behavioral audit, which covers nearly 2,000 scenarios. In a new containment test, it tried to circumvent boundaries about 85% less often than Opus 5. Anthropic also notes the model often suspects it is being evaluated.

Its biology and cyber capabilities are comparable to Claude Mythos 5.1. So Opus 5.5 ships with safeguards similar to Fable 5.1:

• Cybersecurity: Routine bug finding and fixing works. Most other cybersecurity tasks are re-routed to Opus 4.8. The Cyber Verification Program will expand to Opus 5.5.
• Biology: Vetted organizations can apply to the Life Sciences Verification Program.
• Distillation: Preserved thinking stops API users from editing prior context to extract reasoning. It applies to API accounts created on or after August 31, 2026.

Two more changes affect integrations. Thinking can no longer be disabled. Outputs also carry watermarking for EU AI Act compliance. Full details are in the Opus 5.5 System Card.

Interactive Explainer

‘;});tr.innerHTML=h;return tr.children;}
var segOld=mk(document.getElementById(‘tr-old’)), segNew=mk(document.getElementById(‘tr-new’));
function money(v){return ‘$’+(v>=1000?Math.round(v).toLocaleString(‘en-US’):v.toFixed(2));}
function upd(){
var v=ids.map(function(k){var x=+document.getElementById(‘s-‘+k).value;document.getElementById(‘o-‘+k).textContent=x+’M’;return x;});
var a=[v[0]*P5.i,v[1]*P5.o,v[2]*P5.cr,v[3]*P5.cw], b=[v[0]*P55.i,v[1]*P55.o,v[2]*P55.cr,v[3]*P55.cw];
var ta=a.reduce(function(s,x){return s+x},0), tb=b.reduce(function(s,x){return s+x},0);
for(var k=0;k<4;k++){segOld[k].style.width=(ta?a[k]/ta*100:0)+’%’;segNew[k].style.width=(ta?b[k]/ta*100:0)+’%’;}
document.getElementById(‘v-old’).textContent=money(ta);
document.getElementById(‘v-new’).textContent=money(tb);
document.getElementById(‘save’).innerHTML=ta?’Opus 5.5 costs ‘+Math.round((1-tb/ta)*100)+’% less for this token mix.’:’Move a slider to price a workload.’;
}
ids.forEach(function(k){document.getElementById(‘s-‘+k).addEventListener(‘input’,function(){pre.forEach(function(p){p.setAttribute(‘aria-pressed’,’false’)});upd();});});
var pre=[].slice.call(document.querySelectorAll(‘#presets .chip’));
function applyPreset(btn){pre.forEach(function(p){p.setAttribute(‘aria-pressed’,p===btn)});btn.dataset.p.split(‘,’).forEach(function(x,i){document.getElementById(‘s-‘+ids[i]).value=x;});upd();}
pre.forEach(function(b){b.addEventListener(‘click’,function(){applyPreset(b);});});

/* benchmarks */
var M=[[‘Opus 5.5′,’#D97757’],[‘Fable 5.1′,’#3D3D3A’],[‘Opus 5′,’#B0AEA5’],[‘GPT-6 Astra’,’#6A9BCC’],[‘GPT-5.6 Sol’,’#788C5D’]];
var B=[
{n:’Terminal-Bench 4.0′,u:’%’,v:[66.4,55.8,52.3,57.9,37.3],t:’Agentic coding in a terminal. Opus 5.5 reported at xhigh effort, GPT-6 Astra at high effort per OpenAI. At default effort, Opus 5.5 matches GPT-6 Astra at about 40% of the cost.’},
{n:’FrontierCode v1.1′,u:’%’,v:[54.4,50.3,48.0,53.3,47.5],t:’Measures whether an agent’s code changes would be merged. At default (medium) effort Opus 5.5 scores 54.6%, beating GPT-6 Astra’s top 53.3% for about a fifth of the cost per task.’},
{n:’CursorBench 4.0′,u:’%’,v:[57.8,51.8,46.6,null,41.7],t:’Ambiguous multi-file tasks from real Cursor sessions. At default (medium) effort Opus 5.5 scores 52.5%, about 11 points above GPT-5.6 Sol’s top score for about a third of the cost.’},
{n:’GDPval-AA v2.1′,u:’Elo’,v:[1846,1735,1708,1542,1588],min:1200,max:1900,t:’Real-world professional work across 44 occupations, run by Artificial Analysis. Axis starts at 1200 Elo. At default effort, Opus 5.5 beats GPT-6 Astra at max effort for about a fifth of the cost per task.’},
{n:’AutomationBench’,u:’%’,v:[40.0,31.4,26.9,41.4,28.8],t:’Business workflows across connected apps, run by Zapier. GPT-6 Astra leads here. Runs used no fallback models, so safeguard interventions counted as failures for Opus 5.5.’},
{n:’Humanity’s Last Exam’,u:’%’,v:[67.7,65.6,63.6,57.2,null],t:’Multidisciplinary reasoning, with tools.’},
{n:’Terminal-Bench-Science 0.1′,u:’%’,v:[58.7,52.6,29.0,64.6,22.4],t:’Agentic scientific research. GPT-6 Astra leads here. Standard error is 3.5 to 5 points per model.’},
{n:’OSWorld 2.0′,u:’%’,v:[81.8,80.7,74.0,null,null],t:’Computer use, partial-credit scoring.’},
{n:’Chartography’,u:’%’,v:[89.0,88.4,83.4,null,null],t:’Visual chart recognition, with tools.’}
];
var curB=0, bch=document.getElementById(‘bchips’), bars=document.getElementById(‘bars’);
B.forEach(function(b,i){var c=document.createElement(‘button’);c.className=”chip”;c.textContent=b.n;c.setAttribute(‘aria-pressed’,i===0);c.addEventListener(‘click’,function(){curB=i;drawBench(i);});bch.appendChild(c);});
var h=””;M.forEach(function(m,i){h+=”;});bars.innerHTML=h;
function drawBench(i){
var b=B[i], lo=b.min||0, hi=b.max||100;
[].slice.call(bch.children).forEach(function(c,k){c.setAttribute(‘aria-pressed’,k===i)});
var rows=bars.children;
for(var k=0;kCyber Verification Program expands to Opus 5.5 in the coming weeks.’},
{n:’Restricted biology R&D’,to:’B’,h:’Falls back to another model’,p:’Opus 5.5 uses the same biology safeguards as Fable 5.1, which fall back to another model transparently.’,u:’Unlock path: vetted organizations can apply to the Life Sciences Verification Program.’},
{n:’Edit prior context via API’,to:’C’,h:’Preserved thinking stops the edit’,p:’This anti-distillation safeguard stops API users from editing Claude’s prior context to extract its reasoning.’,u:’Applies to API accounts created on or after August 31, 2026. See the preserved thinking docs.’}
];
var curR=0, rch=document.getElementById(‘rchips’), dot=document.getElementById(‘dot’), raf=null;
R.forEach(function(r,i){var c=document.createElement(‘button’);c.className=”chip”;c.textContent=r.n;c.setAttribute(‘aria-pressed’,i===0);c.addEventListener(‘click’,function(){curR=i;runRoute(i);});rch.appendChild(c);});
function setOn(ids){[‘nR’,’nS’,’nA’,’nB’,’nC’,’w0′,’wA’,’wB’,’wC’].forEach(function(id){document.getElementById(id).classList.toggle(‘on’,ids.indexOf(id)>-1);});}
function runRoute(i){
var r=R[i];
[].slice.call(rch.children).forEach(function(c,k){c.setAttribute(‘aria-pressed’,k===i)});
document.getElementById(‘oh’).textContent=r.h;
document.getElementById(‘op’).textContent=r.p;
var ou=document.getElementById(‘ou’);ou.innerHTML=r.u;ou.hidden=!r.u;
var w0=document.getElementById(‘w0’), w1=document.getElementById(‘w’+r.to);
var finalOn=[‘nR’,’nS’,’w0′,’w’+r.to,’n’+r.to];
if(raf)cancelAnimationFrame(raf);
if(reduce){setOn(finalOn);var e=w1.getPointAtLength(w1.getTotalLength());dot.setAttribute(‘cx’,e.x);dot.setAttribute(‘cy’,e.y);dot.setAttribute(‘opacity’,1);setTimeout(resize,60);return;}
setOn([‘nR’]);dot.setAttribute(‘opacity’,1);
var L0=w0.getTotalLength(), L1=w1.getTotalLength(), dur=1300, t0=null;
function step(ts){
if(!t0)t0=ts;var k=Math.min(1,(ts-t0)/dur), d=k*(L0+L1), pt;
if(d<=L0){pt=w0.getPointAtLength(d);if(k>0.05)setOn([‘nR’,’w0′]);}
else{pt=w1.getPointAtLength(d-L0);setOn([‘nR’,’nS’,’w0′,’w’+r.to]);}
dot.setAttribute(‘cx’,pt.x);dot.setAttribute(‘cy’,pt.y);
if(k<1)raf=requestAnimationFrame(step);else setOn(finalOn);
}
raf=requestAnimationFrame(step);
setTimeout(resize,60);
}

applyPreset(pre[0]);
drawBench(0);
window.addEventListener(‘load’,resize);
window.addEventListener(‘resize’,resize);
if(window.ResizeObserver)new ResizeObserver(resize).observe(document.body);
resize();
})();