GPT-6.1Sol
Garaad u dhow Astra oo ku kacaya shan meelood meel qiimaheeda, si joogto loogu helo waxqabadka ugu casriyeysan
Waxaan soo bandhigaynaa GPT‑6.1 Sol, oo ah casriyeyn lagu sameeyey GPT‑6 Sol, lehna waxqabad aad u xooggan oo ku saabsan koodhaynta wakiil-ku-saleysan, adeegsiga kombiyuutarka iyo shaqada xirfadeed. Waxay Sol ku soo dhoweynaysaa GPT‑6 Astra marka la qabanayo hawlaha adag, iyadoo ku kacaysa shan meelood meel qiimayaasha caadiga ah ee token-ka gelinta iyo soo-saarka Astra. Taasi waxay horumariyeyaasha u saamaxaysaa inay isla miisaaniyadda ku dhisaan kuna hawlgeliyaan wakiillo awood badan.
GPT‑6.1 Sol waxay bixisaa waxqabad u dhow Astra oo ku kacaya qiimaha Sol, taas oo ka dhigaysa nooca maanta la heli karo ee ugu kharash-hufan marka loo eego waxqabadkiisa. Gelinta kaydsan waxay ku kacaysaa keliya $0.10 halkii milyan ee token—95% ka jaban qiimaha gelinta caadiga ah—taas oo yaraynaysa kharashka hawlaha wakiillada ee u baahan inay macnaha guud dib ugu adeegsadaan codsiyo soo noqnoqda.
GPT‑6 Astra ayaa guud ahaan weli ah noocayaga ugu awoodda badan ee ugu casriyeysan. GPT‑6.1 Sol waxay bixisaa dheellitir cusub oo awood iyo kharash ah, kuna habboon shaqada muhiimka ah ee isticmaalayaashu rabaan inay marar badan qabtaan iyo macaamiisha API-ga ee dhisa kuna hawlgeliya ablikeeshinno baaxad weyn.
GPT‑6.1 Sol waxay si weyn uga horumartay GPT‑6 Sol xagga shaqada xirfadeed ee adag, laga bilaabo qorista iyo saxidda koodhka ilaa fahamka dukumentiyada iyo fulinta habraacyada ganacsi ee tallaabooyinka badan leh. Dhawr ka mid ah qiimayntan, waxay ku dhowdahay waxqabadka GPT‑6 Astra iyadoo kharashku aad uga yar yahay.
Marka la joogo DeepSWE v1.1, oo qiimaysa hawlaha adag ee injineernimada softiweerka ee kaydadka koodhka dhabta ah, GPT‑6.1 Sol waxay la siman tahay GPT‑6 Astra iyadoo kharashku yahay qiyaastii shan meelood meel. Waxa kale oo ay dhibcaha ugu fiican ee GPT‑6 Sol dhaaftay 6.4 dhibcood oo boqolleey ah, iyadoo dadaalka caqliyeynta iyo kharashkuba ka yar yihiin.
Imtixaanka DeepSWE 1.1(ku furmaa daaqad cusub), wakiillada AI waxay xalliyaan hawlo cusub oo injineernimada barnaamijyada ah oo u baahan shaqo muddo dheer socota.
GDP.pdf waxay cabbirtaa saxnaanta jawaabaha noocyadu ka bixiyaan su'aalaha xirfadeed iyagoo adeegsanaya dukumentiyo PDF oo adag, lehna jadwallo, garaafyo, jaantusyo iyo faahfaahin far yaryar ku qoran. GPT‑6.1 Sol waxay ka dhibco badan tahay Opus 5.5 oo leh beddello gurmad ah, iyadoo kharashka halkii hawl ka yar yahay kala badh dhammaan dejimaha caqliyeynta ee la tijaabiyey. Waxa kale oo ay ku dhowdahay waxqabadka hormuudka ah ee GPT‑6 Astra, iyadoo kharashka halkii hawl yahay qiyaastii shan meelood meel.
Imtixaanka GDP.pdf(ku furmaa daaqad cusub), moodelladu waa inay ka jawaabaan codsiyo ka tarjumaya xaalado dhab ah oo ku saabsan dukumentiyo PDF ah oo kakan, lagana soo qaatay habraacyada shaqo ee maaliyadda, daryeelka caafimaadka, sharciga iyo toddoba qaybood oo kale oo xirfadeed.
AutomationBench, oo cabbirta in wakiilladu si sax ah u dhammaystiraan habraacyada ganacsi ee tallaabooyinka badan leh, GPT‑6.1 Sol waxay Opus 5.5 kaga sarraysaa 2.2 dhibcood oo boqolleey ah marka dadaalka caqliyeynta la dhigo heer Dhexdhexaad ah, iyadoo kharashku yahay qiyaastii saddex meelood meel. Dhibcahaasi waxay sidoo kale 4.8 dhibcood oo boqolleey ah ka sarreeyaan kuwa GPT‑6 Sol ee isla dejintaas.
In AutomationBench 1.0.6(ku furmaa daaqad cusub), AI agents are tested on end-to-end workflows using 47 tools across sales, marketing, operations, support, finance, and HR. The datapoint for Claude Fable 5.1 understates its actual cost, as it omits the cost of fallbacks, which occurred on ~40% of tasks.
GPT‑6.1 Sol waxay sidoo kale horumar weyn ka samaysay hawlaha u baahan la falgalka ablikeeshinnada kombiyuutarka. Qaybta aan internetka ku xidhnayn ee OSWorld 2.0, oo wakiillada ku qiimaysa habraacyada adag ee adeegsiga kombiyuutarka, GPT‑6.1 Sol waxay GPT‑6 Sol kaga sarraysaa toddoba dhibcood oo boqolleey ah marka dadaalka caqliyeyntu ugu sarreeyo, iyadoo kharashku ka yar yahay kala badh. Waxay dhibcaha Astra u soo jirsataa 2.1 dhibcood oo boqolleey ah marka dadaalka caqliyeyntu ugu sarreeyo, iyadoo kharashka halkii hawl yahay qiyaastii toddoba meelood meel.
In OSWorld 2.0(ku furmaa daaqad cusub), AI agents attempt long-horizon computer-use workflows spanning everyday and professional tasks. We report the partial reward on the offline set from the v2026.08.08 release.
Terminal-Bench Science 0.1, oo qiimaysa habraacyada sayniska sida falanqaynta xogta, jilidda iyo caddaynta aragtiyaha xisaabta, GPT‑6.1 Sol waxay heshaa in ka badan laba jibbaarka dhibcaha GPT‑6 Sol marka dadaalka caqliyeyntu ugu sarreeyo, iyadoo kharashka halkii hawl ka yar yahay kala badh. Marka dadaalku ugu sarreeyo, GPT‑6.1 Sol waxay celcelis ahaan ku kacaysaa $5.47 halkii hawl, marka loo eego $23.21 Opus 5.5 iyo $23.80 Astra. Waxay sidaas ku bixisaa awood saynis oo xooggan, iyadoo mid kasta oo labadaas nooc ka jaban wax ka badan 75%.
GPT‑6 Astra ayaa weli ugu dhibco badan noocyada la tijaabiyey, iyadoo heshay 68.1%, waana in loo adeegsadaa hawlaha cilmi-baarista sayniska ee ugu adag.
Imtixaanka Terminal-Bench Science 0.1(ku furmaa daaqad cusub), wakiilladu waxay dhammaystiraan habraacyada cilmi-baarista sayniska iyagoo adeegsanaya koodh iyo qalabka terminalka, oo ay ku jiraan falanqaynta xogta, fulinta jilitaannada iyo waafajinta moodellada xogta.
GPT‑6.1 Sol waxay sidoo kale hagaajisaa saxnaanta xaqiiqooyinka marka codsiyadu adag yihiin. Marka dadaalka caqliyeyntu aad u sarreeyo, heerkeeda khaladaadka xaqiiqadu waa 4.1%, kana hooseeya 4.5% ee GPT‑6 Sol, kuna dhow 4.0% ee Astra isla dejintaas. Haddana kharashka halkii hawl wuxuu Astra ka yar yahay qiyaastii 83%.
Qiimayntani waxay cabbirtaa saamiga jawaabaha leh ugu yaraan hal khalad xaqiiqo, iyadoo adeegsanaysa wada sheekaysiyo xogta aqoonsiga laga saaray oo isticmaalayaashu ku calaamadiyeen khalad uu sameeyey nooc hore. Codsiyadan si ula kac ah loo adkeeyey ma matalaan isticmaalka caadiga ah.
Waxaan saxnaanta xaqiiqooyinka ku qiimaynaa wada sheekaysiyo ChatGPT oo xogta aqoonsiga laga saaray, kuwaas oo isticmaalayaashu ku calaamadiyeen khalad xaqiiqo oo nooc hore sameeyey. Wada sheekaysiyadan khaladaadka dhaliya ma matalaan isticmaalka caadiga ah, oo khaladaadka xaqiiqadu ku yar yihiin.
Qiimayntayada waafaqsanaanta, GPT‑6.1 Sol waxay muujisaa horumar weyn marka loo eego GPT‑6 Sol, taas oo ku soo dhoweynaysa GPT‑6 Astra.
GPT‑6.1 Sol waxay si ka hufan u sheegtaa xaddidaadaheeda, waxayna ku kalsooni badan tahay dhowrista ujeeddada isticmaalaha iyo xaddidaadaha badbaadada. Qiimaynaha adag, heerarkeeda fashilku way ka hooseeyaan kuwa GPT‑6 Sol xagga sheegidda ciladaha agabka raadinta, dhowrista xaddidaadaha cad iyo ka fogaanshaha natiijooyinka aan la oggolaan inta lagu jiro hawlaha wakiillada. Ma aannu arag wax isku dayo ah oo lagu hareer marayo dib-u-eegaha badbaadada ee otomaatigga ah, si la mid ah GPT‑6 Astra iyo GPT‑6 Sol.
Qiimaynaha hoose waxay si ula kac ah u tijaabiyaan xaalado adag, mana cabbiraan heerarka fashilka ee isticmaalka caadiga ah.
Laga bilaabo maanta, GPT‑6.1 Sol waxaa ChatGPT Shaqo iyo Codex ka heli kara dhammaan isticmaalayaasha Plus, Pro, Business, Enterprise iyo Edu. Noocyadan weli lagama heli karo Sheekaysi. Horumariyeyaashu waxay sidoo kale API-ga OpenAI kaga heli karaan magaca gpt-6.1-sol. Qiimayaasheeda API-ga ee caadiga ahi waa $2 halkii milyan ee token-ka gelinta, $0.10 halkii milyan ee token-ka gelinta kaydsan iyo $10 halkii milyan ee token-ka soo-saarka.
Waxaan sidoo kale daahfuraynaa GPT‑6 Astra Ultrafast, oo ah nooca ugu casriyeysan ee dunida ugu dheereeya, kana xawaare badan GPT‑6 Astra ilaa 8 jeer, iyo weliba GPT‑6.1 Sol Ultrafast. Kuwan waxaa laga heli karaa API-ga, sidoo kalena ChatGPT Shaqo iyo Codex ayay ka heli karaan isticmaalayaasha heerkeenna cusub ee Pro oo leh xaddiga isticmaalka ugu badan, qiimihiisuna yahay $500 bishii. GPT‑6.1 Sol Ultrafast waxay horumariyeyaasha siinaysaa garaad u dhow Astra oo xawaarihiisu gaarayo ilaa 8 jibbaar, iyadoo kharashka API-gu qiyaastii la mid yahay kii hore ee GPT‑6 Astra.
Qorayaa
Evaluations of GPT were performed in our research environment or via our API, which may provide slightly different output from production ChatGPT due to differences in the system prompts, tools available, efforts, etc. Evaluations of competitor models were taken from publicly available reports.
