OpenAI has released GPT-6.2 Astra and GPT-6.2 Sol, twp closed models aimed at professional and agentic work. The GPT-6.2 Astra model is the new frontier tier for hard, long-horizon tasks. GPT-6.2 Sol is the efficient everyday tier. For developers, the key point is cost per task. Astra’s token prices fall 20%, and OpenAI says both models finish tasks with fewer output tokens. Both are live in the OpenAI API as gpt-6.2-astra and gpt-6.2-sol.
TL;DR
- Size: 1,050,000-token context window, 922,000 max input, 128,000 max output.
- Runs on: OpenAI’s hosted API, ChatGPT Work and Codex only.
- Performance: Astra scores 62 on the Artificial Analysis Intelligence Index, second behind Claude Fable 5.2 (65).
- Best: Astra ranks #1 on AA-3D with 1,318 Elo, ahead of Fable 5.2 (1,271).
- Worst: Astra posts a 59% AA-Omniscience hallucination rate (Sol: 55%). Sol’s Coding Agent Index drops 3 points.
- Bottom line: Best: frontier results at much lower cost per task. Worst: no coding gains, and hallucination stays high.
What is GPT-6.2 Astra?
GPT-6.2 Astra is OpenAI’s frontier reasoning model for complex professional work. It succeeds GPT-6 Astra. OpenAI says it produces cleaner slides, spreadsheets and documents that follow templates and field conventions. It also builds better frontends, games and 3D scenes.
GPT-6.2 Sol replaces GPT-6.1 Sol as the everyday model.
How do the two models differ?
Both accept text and image input and return text. Both share the same context limits. The split is cost and depth.
- Astra: knowledge cutoff May 18, 2026. Built for the hardest multistep work.
- Sol: knowledge cutoff June 1, 2026. Built for fast, versatile everyday tasks.
Neither model accepts reasoning effort none. OpenAI recommends keeping prompt prefixes unchanged and appending new instructions. That improves cache reuse and preserves reasoning across turns.
What do the benchmarks show?
OpenAI’s own claims focus on cost-adjusted gains:
- Agents’ Last Exam v1.1: Astra scores 5.2 points above GPT-6 Astra at under a third of the cost. It leads Claude Fable 5.2 by 5.5 points at 5% of the cost.
- Terminal-Bench-Science 0.1: Astra gains 9 points over GPT-6 Astra at under a third of the cost.
- GDPval v2 (internal): Astra gains 6.2 points with 22% fewer output tokens. Sol gains 5.3 points over GPT-6.1 Sol.
- FrontierCode 2.0: Sol lands within 0.5 points of Astra at about one-fifth the cost.
- AA-AnalystAgent: Astra answers 62.5% of questions correctly on all 5 attempts.
Independent results from Artificial Analysis mostly hold up. Astra scores 62 at max effort, with a $4.00 cost per Intelligence Index task. Fable 5.2 scores 65, Gemini 4 Carbon 60, Claude Opus 5.5 59 and Sol 57. Sol matches GPT-6 Astra’s score at about 15% of its cost per task.
On Harvey LAB-AA, Astra passes 15.0% of legal assignments. The Astra for Law configuration passes 20.8%. Epoch’s Capabilities Index puts Astra at 174, tied with Fable 5.2.
‘;});
document.getElementById(‘mtp-bars’).innerHTML=h;document.getElementById(‘mtp-bnote’).textContent=d.note;
setTimeout(function(){R.querySelectorAll(‘.fl’).forEach(function(f){f.style.width=f.dataset.w+’%’;});},40);setTimeout(post,60);}
R.querySelectorAll(‘.seg button’).forEach(function(b){b.addEventListener(‘click’,function(){R.querySelectorAll(‘.seg button’).forEach(function(x){x.classList.remove(‘on’)});b.classList.add(‘on’);draw(b.dataset.b);});});
function fmt(n){return ‘$’+n.toLocaleString(‘en-US’,{minimumFractionDigits:2,maximumFractionDigits:2});}
function calc(){var i=+document.getElementById(‘mtp-in’).value,o=+document.getElementById(‘mtp-out’).value,q=+document.getElementById(‘mtp-rq’).value;
document.getElementById(‘mtp-in-l’).textContent=i.toLocaleString(‘en-US’);document.getElementById(‘mtp-out-l’).textContent=o.toLocaleString(‘en-US’);document.getElementById(‘mtp-rq-l’).textContent=q.toLocaleString(‘en-US’);
var lc=i>272000,mi=lc?2:1,mo=lc?1.5:1;
var a=(i*8*mi+o*40*mo)/1e6*q,s=(i*2*mi+o*10*mo)/1e6*q;
document.getElementById(‘mtp-ca’).textContent=fmt(a);document.getElementById(‘mtp-cs’).textContent=fmt(s);
document.getElementById(‘mtp-warn’).textContent=lc?’Long-context tier applied: input over 272K tokens.’:”;}
[‘mtp-in’,’mtp-out’,’mtp-rq’].forEach(function(id){document.getElementById(id).addEventListener(‘input’,calc);});calc();
var S=[[‘1. Everyday agent or chat workload?’,’Start with GPT-6.2 Sol. It costs $2/$10 per 1M tokens and scores 57 on the AA Intelligence Index.’],
[‘2. Code changes to real repos?’,’Sol lands within 0.5 points of Astra on FrontierCode 2.0 at about one-fifth the cost. Test Sol first.’],
[‘3. Hard, long-horizon professional work?’,’Use GPT-6.2 Astra for legal, finance, science and CAD-heavy jobs. It leads Agentsx27 Last Exam per OpenAI.’],
[‘4. Keep costs predictable’,’Keep prompt prefixes unchanged for cache reuse. Stay under 272K input tokens to avoid long-context rates.’]];
var si=0;function step(){document.getElementById(‘mtp-sh’).textContent=S[si][0];document.getElementById(‘mtp-sp’).textContent=S[si][1];
R.querySelectorAll(‘.steps i’).forEach(function(x,j){x.classList.toggle(‘on’,j<=si)});
document.getElementById(‘mtp-prev’).disabled=si===0;document.getElementById(‘mtp-next’).disabled=si===S.length-1;setTimeout(post,60);}
document.getElementById(‘mtp-prev’).addEventListener(‘click’,function(){if(si>0){si–;step();}});
document.getElementById(‘mtp-next’).addEventListener(‘click’,function(){if(si


