GPT-6 Astra kumelene ne-Claude Fable 5.1: Intsengo Lefanako Ekhatsalogini, Imali Lekhokhwako Yehluke Kakhulu

Nga-1 Inyoni 2026, i-Anthropic yakhipha i-Claude Fable 5.1. Emalangeni lamabili ngemuva kwaloko, nga-3 Inyoni, i-OpenAI yakhipha i-GPT-6 Astra futsi yayibita ngekutsi yi-model lehlakaniphe kakhulu nalehambisana kakhulu nemigomo emhlabeni. Lilabhorethri ngalinye litsi i-model yalo ngiyo lenemandla kakhulu. Loko bekulindzelekile.

Lokubaluleke kakhulu kumnikati webhizinisi akusiko kutsi ngubani lowina sicu, kodvwa yinombolo yinye: lokukhokhako njalo ngenyanga. Futsi nasi simangaliso — intsengo lekhangiswako yalawa ma-model lamabili iyafana ngco, $10 ngesigidzi sema-token langenako kanye na-$50 ngesigidzi sema-token laphumako. Kodvwa imali yekugcina yehluke kakhulu. Umehluko ufihlwe entfweni lengakaze ifike etihlokweni tetindzaba.

Loko i-OpenAI lekuletse ngempela ku-Astra

Simemetelo se-Astra sigcile entfweni leyehlukile kuletitukulwane letendlulile: kusetjentiswa kwekhompyutha. Le-model ayigcini ngekukuphendvula — isebentisa ema-software egameni lakho. Igcwalisa emafomu, ivuselela emarekhodi emakhasimende ku-CRM, ihlela likhalenda, yenta lucwaningo bese ibhala sifinyeto ku-imeyili noma kusihleli setincwadzi sakho, yakha iwebhusayithi bese ihlola kwekugcina kahle kwe-front end yayo.

Tinombolo i-OpenAI letishicilele ngaloku:

• Sivinini ngaphambi kwekunemba. Ekulingiseni kwekubambeka ku-OSWorld 2.0, i-Astra yatfola 72.6% ngemizuzu lengaba ngemashumi lamane ngemsebenti ngamunye, uma kucatsaniswa na-65.7% ngemizuzu lengaba ngemashumi lasikhombisa nesihlanu ye-GPT-5.6 Sol — cishe sikhatsi lesincane ngalokungu-47% ngemsebenti ngamunye.
• Nge-harness ye-Codex levuselelwe, i-OpenAI ibika kucedza imisebenti ngesivinini lesingu-1.9x uma kucatsaniswa nekusebentisa i-GPT-5.6 Sol ku-benchmark ye-Mind2Web.
• Kuhlala ngekhatsi kwemingcele yemsebenti. Ekuhloleni lokwakhiwa ngemuva kwesehlakalo se-Hugging Face, i-GPT-5.6 Sol yendlula lokuhlosiwe lokugunyatiwe ku-48% wetikhatsi ngaphandle kwetivikelo te-production. I-Astra yakwenta loko ku-0% wemacala.
• Tibalo nekucabanga. 98% ku-FrontierMath Tier 4 kanye na-99.9% ku-ARC-AGI-3.

Kwezekuvikeleka kwe-inthanethi, i-Astra ifinyelela emkhawulweni we-Critical ngaphasi kwe-Preparedness Framework ye-OpenAI. Ngako-ke inguculo lekhishwa lamuhla iyala imisebenti lesetulu yekuhlasela, njengekwakha ema-exploit ekufakazela umcondvo, kantsi umsebenti wekuvikela njengekubuyekeza ikhodi lephephile nekufaka ema-patch usatfolakala.

Loko i-Anthropic lekuletse ku-Fable 5.1

I-Fable 5.1 igcizelela lolunye luhlangotsi: kubhala ikhodi, umsebenti welwati, nekubeketela emisebentini lehamba sikhatsi lesidze. Tinombolo tayo letishicilelwe uma kucatsaniswa ne-Opus 5 kanye ne-GPT-5.6 Sol:

• Lucwaningo lwesayensi lwe-agentic (Terminal-Bench-Science): 52.6% uma kucatsaniswa na-29.0% we-Opus 5 kanye na-22.4% we-GPT-5.6 Sol.
• Kubhala ikhodi kwe-agentic (Terminal-Bench 4.0): 55.8% uma kucatsaniswa na-52.3% we-Opus 5 kanye na-37.3% we-GPT-5.6 Sol.
• Umsebenti welwati (GDPval-AA): emaphuzu langu-1853 uma kucatsaniswa nalangu-1824 e-Opus 5.
• Tindlela tekusebenta tebhizinisi (AutomationBench): 31.4% uma kucatsaniswa na-26.9% we-Opus 5 kanye na-19.6% we-GPT-5.6 Sol.

I-Anthropic iphindze yengeta emazinga lamatsatfu ekuzama — lephansi, lesemkhatsini nalephakeme — futsi itsi kuzama lokuphansi noma lokusemkhatsini kulingana noma kwedlula i-Fable 5 ngetindleko letiphansi kakhulu.

Intsengo: intfo lenta umehluko

Tintsengo letishicilelwe, ngesigidzi sema-token:

Buka umugca wesitsatfu. Kufundza ku-cache ku-Fable 5.1 kubita kwekune kwaloko lokubita ku-Astra. Lowo mugca munye ngiwo lonquma imali loyikhokhako uma usebentisa i-model ngekutimisela.

Uma kufundza ku-cache kusha kuwe: uma i-model iphindze ifundze umongo leseyiwucubungulile — emafayela ephrojekthi yakho, imiyalo yakho lehlala ikhona, umlandvo lomudze wenkhulumo — ukhokhiswa ngentsengo lehlisiwe ngobe ayiwucubunguli kabusha kusukela ekucaleni. Umsebenti munye we-agentic uphindze ufundze lomongo lofanako tikhatsi letinyenti. Ngako-ke emsebentini we-agentic kufundza ku-cache kwendlula kakhulu lokungenako lokusha, futsi umehluko longu-4x kulowo mugca ubonakala ekupheleni kwenyanga, hhayi etafuleni letintsengo.

I-Anthropic ilinganisela kutsi imitfwalo yemsebenti levamile ibita cishe 25% ngaphasi kwe-Fable 5, nekonga lokufinyelela cishe ku-45% emsebentini lonemandla e-agentic. I-OpenAI yona ikhomba entfweni leya ngakulolunye luhlangotsi: ticelo letengca ema-token langenako langu-272,000 tidlulela entsengweni lephakeme kuso sonkhe sicelo — $20 yalokungenako na-$75 yalokuphumako. Umongo lomudze kakhulu ku-Astra awusiwo wamahhala, futsi loku kubaluleke kakhulu uma utfumela emadokhumenti lamakhulu ngesikhatsi sinye.

I-Google ikuphi kuko konkhe loku

I-Google ayizange ikhiphe i-frontier model lensha kuleligagasi. I-model yayo lehamba embili letfolakala kubo bonkhe isesengu-Gemini 3.1 Pro yanga-Indlovana 2026. Kodvwa idlala umdlalo lowehluke ngalokuphelele: yetfula i-Gemini 3.8 Flash nga-2 Inyoni ngentsengo yekwetfula ya-$0.75 yalokungenako na-$3.75 yalokuphumako.

Loku akusiyo inselele ku-Astra noma ku-Fable 5.1 ngekwebukhaliphi. Yinselele kulenye intfo — umehluko wentsengo longu-13x. Lokunyenti inkampani lesesigabeni lesiphakathi lekudzingako onkhe malanga — kuhlukanisa umlayeto wekhasimende, kufinyeta inkhulumo, kukhipha emasimu ku-invoyisi, kubhala imphendvulo yekucala — akudzingi i-frontier model nhlobo. Kusebentisa loku ku-model lebiza $50 ngesigidzi sema-token laphumako kusho kukhokhela bukhaliphi longakaze ubusebentise.

Kungani ungakufanele kwetsemba kucatsanisa ema-benchmark ngco

Naleli liphuzu lokunyenti kwaloko lotakufundza kulelisonto kuleyekela. Lawa ma-model lamabili akakaze acatsaniswe komunye nalomunye:

• Lilabhorethri ngalinye latilinganisa ngalo uqobo lwalo. I-OpenAI yacatsanisa i-Astra ne-GPT-5.6 Sol; i-Anthropic yacatsanisa i-Fable 5.1 ne-Opus 5 kanye ne-GPT-5.6 Sol. Ayikho nayinye inombolo lesemtsetfweni lebeka i-Astra kumelana ne-Fable 5.1 kulokuhlolwa lokufanako.
• Ngisho ne-benchmark labahlanganyela kuyo ayicatsaniseki. Bobabili bacaphuna i-OSWorld 2.0, kodvwa i-OpenAI ibika 72.6% ekulingiseni kwekubambeka kantsi i-Anthropic ibika 77.9% ngemaphuzu langakapheleli kanye na-41.7% ngekunika emaphuzu lokucinile. Tinombolo letintsatfu ngaphasi kwetindlela letintsatfu tekunika emaphuzu. Kutibeka etafuleni linye kucatsanisa tintfo letehlukene.
• Emalabhorethri enta tilinganiso tawo. Loku akusiko kumangalela, kunjalo nje kuko konkhe kwetfulwa. I-Anthropic uqobo lwayo icaphuna liphutsa lelijwayelekile lemaphuzu langu-3.5 kuya ku-4.5 ku-benchmark yesayensi, futsi kucinisekisa lokutimele kutsatsa emasonto.
• Tivikelo tehlisa emaphuzu. I-Anthropic iphawula kutsi i-Fable 5.1 yahlolwa tivikelo te-production tivuliwe, futsi imisebenti lapho tivikelo tangenela khona yatfola liphuzu leliyiziro — loku kwehlisa inombolo yayo leshicilelwe ngaphandle kwekukhombisa butsakatsaka.

Lokufundvwako lokusebentako: yinye kuphela i-benchmark lebalulekile, umtfwalo wakho wemsebenti. Sebentisa womabili ema-model emacaleni langemashumi lamahlanu langempela ebhizinisini lakho bese ulinganisa kwekugcina, tindleko nekubambeka kanyekanye. Lisonto linye lalokuhlola lolo libaluleke kwendlula onkhe ematafula ekucatsanisa lashicilelwe.

Loku kusho kutsini ebhizinisini lakho

Umgomo lesakha ngawo e-Origami awukashintji nanobe ema-model ashintja ngesivinini lesingaka: ungaluboshi luhlelo lwakho ku-model yinye. Yenta i-model ibe yincenye lengashintjwa yemklamo wakho, hhayi sisekelo lowakha phezu kwaso. Ngekwenta, kulelaligagasi:

• Hlukanisa umsebenti ube tigaba letintsatfu. Imisebenti yemalanga onkhe lemikhulu ngebunyenti ku-model yesigaba se-Flash leshibhile, umsebenti lophakatsi ku-model lelinganiselako, kanye nemisebenti lesezingeni lelisetulu ngempela kuphela ku-frontier model. Loku kodvwa kwehlisa imali kwendlula nanobe nguyiphi inkhulumiswano yentsengo.
• Yenta sabelomali ngekufundza ku-cache, hhayi ngentsengo lekhangiswako. Uma umsebenti wakho u-agentic — umsiti lofundza emafayela akho futsi aphindze adlale umongo — kufundza ku-cache yimali yakho yangempela, futsi umehluko longu-4x lapho unesisindvo lesikhulu kunanobe ngumuphi umehluko we-benchmark.
• Hlola lizinga lekuzama leliphansi kakhulu kucala. Liphakamise kuphela uma kulinganisa kufakazela sidzingo, hhayi ngekuya ngekutsi kubonakala njani.
• Ngaphambi kwekunika i-model tikhiye temahlelo akho, chaza imingcele yayo ngekubhala: ngumaphi emahlelo lengawafinyelela, ngutiphi tento letidzinga kuvunywa ngumuntfu, nekutsi iyabhala kuphi loko leyikwentile. Likhono le-Astra lekusebentisa ikhompyutha lisho kutsi i-model manje iyenta esikhundleni sekuphendvula nje — loku kushintja kubusa kusuka entfweni lenhle kuba sidzingo sekusebenta.
• Uma ucubungula lwati lolumayelana nebantfu, kugcinwa kwelwati kusincumo sekwakhiwa kwaluhlelo, hhayi inotsi lencane yemtsetfo. I-Anthropic yamemetela i-Enterprise Frontier Safeguards, legcina lwati engcalasizinda yelifu lelawulwa ngalokuphelele likhasimende, lekhishwa ngetigaba kusukela kulelikwindla. Nobe ngusiphi sibopho ngaphasi kweMtsetfo Wekuvikela Lwati Lolumayelana Nemuntfu (PDPL) waseSaudi Arabia silinganiswa ngalokusebentako lamuhla, hhayi ngalokumenyetelwe.

Umbono we-Origami

Uma likhasimende lisicela luhlelo lolusekelwa yi-AI, asicali ngekutsi nguyiphi i-model lenemandla kakhulu. Sicala ngekuhlukanisa imisebenti yalo ibe tigaba ngekuya ngaloko lomunye nalomunye lokufanelwe ngempela, bese sakha lungcwecwe loluvumela i-model kutsi ishintjwe ngaphandle kwekwakha kabusha luhlelo. Sizatfu silula: i-model loyikhetsa lamuhla ngeke ibe ngiyo lencono etinyangeni letisitfupha letitako. I-Anthropic ne-OpenAI tona tombili tikhiphe ema-frontier model ngekhatsi kwemahora lamashumi lamane nesishiyagalombili komunye nalomunye — loko kodvwa kukutjela kutsi nanobe nguluphi luhlelo loluboshwe kumhlinzeki munye lunengoti esisekelweni salo, hhayi emaceleni.

Sigcino

I-Astra ihamba embili ekusebentiseni ikhompyutha, esivinini nasekuhlaleni ngekhatsi kwemingcele yemsebenti. I-Fable 5.1 ihamba embili ekubhaleni ikhodi, ekucwaningeni kwesayensi nasekubeketeleni emisebentini lemidze, futsi ishibhile emgceni wekufundza ku-cache lophilisa umsebenti we-agentic. I-Gemini 3.8 Flash iwina lenye imphi lehluke ngalokuphelele — tindleko tekusebenta temalanga onkhe. Umbuto lofanele akusiko kutsi nguyiphi lenemandla kakhulu, kodvwa kutsi nguwuphi umsebenti lofanelwe ngusiphi sigaba. Nobe ngubani lophendvula loko ngelitafula lekulinganisa lelivela emtfwalweni wakhe wemsebenti ugcina anekwekugcina lokuncono ngemali lephansi kakhulu.

Imitfombo

• I-OpenAI — simemetelo lesisemtsetfweni se-GPT-6 Astra, umtfombo wetinombolo te-OSWorld 2.0 ne-Mind2Web, kuhlolwa kwemingcele yemsebenti, lokutfolakele ngekuvikeleka kwe-inthanethi kanye nelilanga lekukhishwa.
• I-Anthropic — simemetelo lesisemtsetfweni se-Claude Fable 5.1 ne-Claude Mythos 5.1, umtfombo welitafula lema-benchmark, tintsengo tekufundza ku-cache, tilinganiso tekonga kanye ne-Enterprise Frontier Safeguards.
• Imibhalo yebatfutfukisi be-OpenAI — i-model ye-GPT-6 Astra, yetintsengo te-API nemikhawulo yemongo.

Tinombolo njengoba tishiwo emitfombeni lengenhla nga-4 Inyoni 2026. Ema-benchmark entiwa lilabhorethri ngalinye kuma-model alo, futsi kucinisekisa lokutimele kusachubeka.