GPT-6 Astra: Kizazi kipya cha kiwango cha ufikiri

GPT

Astra

Kizazi kipya cha Kiwango cha ufikiri

Tunawaletea GPT‑6 Astra, muundo wenye akili zaidi na unaolingana zaidi duniani.

GPT‑6 Astra inaunganisha miaka ya utafiti na uwekezaji mkubwa katika mafunzo ya awali, mafunzo ya uimarishaji na ulinganishaji. Astra ni ya hali ya juu katika matumizi ya kompyuta, kuvinjari mtandaoni, uhandisi wa programu, usalama wa mtandao, sayansi na kazi za kitaalamu. Astra inafikia ukomo wa FrontierMath Tier 4 kwa alama ya asilimia 98, ikiwa tayari imesaidia kutatua matatizo ya muda mrefu ambayo hayajatatuliwa⁠ katika hisabati. Astra pia imefikia kiwango cha juu zaidi kwenye ARC-AGI-3 kwa alama ya asilimia 99.9 na kwenye ExploitBench kwa alama ya asilimia 100. Pia inaweka mpaka mpya katika matumizi ya kompyuta na kivinjari, ikishughulikia kazi za kitaalamu zenye mahitaji makubwa zaidi kwa kasi, usahihi na uamuzi usio na kifani.

GPT‑6 Astra inatolewa leo hatua kwa hatua kwa idadi ndogo ya mashirika na, katika siku zijazo, itapatikana kwa watumiaji wote wa ChatGPT Plus, Pro, Business na Enterprise, pamoja na kupitia OpenAI API, Microsoft Azure na AWS Bedrock.

Terminal-Bench Science 0.1 hupima ikiwa mawakala wanaweza kukamilisha taratibu za kazi za utafiti wa kisayansi kwa kutumia msimbo na zana za taminali, ikijumuisha kuchanganua data, kuendesha uigaji, na kutosheza miundo. GPT‑6 Astra yafikia kiwango kipya cha juu miongoni mwa miundo iliyolinganishwa kwa asilimia 64.6, dhidi ya asilimia 52.6 ya Claude Fable 5.1, kwa gharama ya API iliyokadiriwa kuwa chini kwa takriban asilimia 31. Katika mpangilio wa gharama ya chini, Astra yapata asilimia 61.1, dhidi ya matokeo bora zaidi ya GPT‑5.6 Sol ya asilimia 22.4, kwa gharama ya API iliyokadiriwa kuwa chini kwa takriban asilimia 27.

“Kwenye ARC-AGI-3, Astra ilizidi kiwango chetu cha msingi cha ufanisi wa vitendo vya binadamu katika asilimia 96 ya viwango, na hivyo kufikia kwa ufanisi usawa na binadamu kwenye kigezo hicho. Huu sio tu muundo bora zaidi ambao tumewahi kuujaribu, bali pia unawakilisha hatua kubwa ya mabadiliko katika utendaji wa miundo ya frontier - si tu katika uwezo wake wa kupitia na kutatua mazingira mapya, bali pia katika jinsi inavyojifunza kufanya hivyo kwa ufanisi.”

Greg Kamradt, ARC Prize Foundation

Astra ni muundo wetu uliolinganishwa zaidi, wenye maboresho makubwa katika kuelewa dhamira ya mtumiaji na tabia ya muundo—unaweza kukabidhi kazi kwa kujiamini zaidi katika uamuzi wa Astra. Kama njia moja ya kupima hili, tulitengeneza tathmini mpya iliyoongozwa na tukio la Hugging Face, inayotathmini ikiwa muundo unaokabiliwa na kazi ngumu au isiyowezekana utavuka upeo wake uliokusudiwa. Ikilinganishwa na GPT‑5.6 Sol, ambayo bila ya hatua za ulinzi katika uzalishaji ilivuka lengo lililoidhinishwa katika asilimia 48 ya matukio, GPT‑6 Astra ilifanya hivyo katika asilimia 0 ya matukio.

Muundo bora zaidi duniani wa kutumia kompyuta

GPT‑6 Astra inaashiria mpaka mpya katika kasi, usahihi na usalama wa matumizi ya kompyuta. Inaweza kushughulikia kazi za kuchosha kama kujaza fomu za mtandaoni, kusasisha rekodi za wateja katika CRM, na kupanga kalenda yako. Inaweza kufanya utafiti mtandaoni na kuandaa mihtasari ya awali katika barua pepe yako au katika kihariri chako cha hati. Inaweza kuchanganua data za kisayansi, kuzalisha michoro, kuunda tovuti, na kuendesha ukaguzi wa QA wa sehemu ya mbele ili kuhakikisha kwamba vipengele vyote kwenye tovuti hiyo vinafanya kazi. Inaweza kukusaidia kusakinisha na kujaribu programu kiotomatiki, na kutatua matatizo unayoyaona kwenye skrini. Maboresho haya pia yanaonekana katika matokeo yetu ya tathmini ya kisasa zaidi.

Mtihani wa Mwisho wa Mawakala huwapima mawakala katika majukumu changamano ya kitaalamu ndani ya programu halisi, kuanzia uundaji wa miundo ya kifedha hadi uhandisi na utayarishaji wa maudhui ya vyombo vya habari. GPT‑6 Astra inafikia kiwango kipya cha juu katika ulinganisho unaoonyeshwa, ikipata asilimia 59.3, ikilinganishwa na asilimia 55.5 kwa Claude Opus 5 na asilimia 53.6 kwa GPT‑5.6 Sol. Katika mipangilio hii yenye alama za juu zaidi, Astra pia hutumia takriban tokeni chache kwa asilimia 65 kuliko Opus 5.

Maboresho haya pia husababisha ongezeko kubwa la ufanisi katika kazi halisi za maarifa. Katika uigaji wa muda wa kusubiri kwenye OSWorld 2.0, Astra inafikia utendaji wa juu zaidi wa matumizi ya kompyuta kwa kutumia muda mfupi kwa takribani asilimia 47 kwa kila jukumu kuliko GPT‑5.6 Sol, ikiwa na alama ya asilimia 72.6 kwa takriban dakika 40 kwa kila kazi, ikilinganishwa na asilimia 65.7 kwa takriban dakika 75.^³

Uwezo wa GPT‑6 Astra wa kutumia kompyuta unaweza kuonekana katika matokeo kwenye nyanja mbalimbali, zikiwemo ukuzaji wa michezo, uhandisi wa umeme, na kazi za kila siku zinazohitaji maarifa:

Huu ni uchezaji uliofupishwa wa sekunde 15 wa GPT‑6 Astra ikifanya mpangilio wa bodi ya mzunguko iliyochapishwa (PCB) katika KiCad, ikibadilisha mchoro wa mzunguko wa kielektroniki kuwa PCB inayoweza kutengenezwa kwa kuweka vipengele na kuelekeza miunganisho ya shaba. Kwa kuwa ni sehemu muhimu ya kila kifaa cha kielektroniki leo, mpangilio wa PCB ni kazi inayofanywa kwa mikono na ni chanzo cha kawaida cha ucheleweshaji katika mchakato wa kubuni vifaa vya kielektroniki. Kuharakisha kazi hii kunawaweka huru wahandisi ili wavumbue, waboreshe na kujaribu wazo lao linalofuata kwa kasi ya juu zaidi kwa kiasi kikubwa.

Alongside Astra, we are also updating the Codex harness to significantly improve the speed of computer use. Combined with Astra’s efficiency, this translates to a 1.9x faster task completion compared to the current GPT‑5.6 Sol experience, on the Mind2Web benchmark. The model’s improvements on speed mean it can take on many time-consuming life tasks for you, faster than you can.^⁴

GPT‑6 Astra: dakika 2 sekunde 54

“Tunaunganisha GPT‑6 Astra katika mazingira ya uendeshaji ya Devin siku ya uzinduzi, ambapo inatoa utendaji wa kisasa zaidi kwenye kipimo chetu cha ndani cha majaribio. Matumizi yake bora ya kompyuta, uandishi na uelewa wa msingi wa msimbo yaliboresha majaribio moja kwa moja: video ni rahisi zaidi kufuatilia, na ripoti ni wazi zaidi na fupi zaidi”

Silas Alberti, Makamu wa Rais Mwandamizi wa Utafiti, Cognition

Mabadiliko makubwa katika kazi za kitaalamu

GPT‑6 Astra inaunganisha maendeleo katika matumizi ya kompyuta na mafunzo yanayolengwa kwa mazingira ya kitaalamu, ili kusaidia kushughulikia majukumu changamano ya kazi. Inachanganya Kiwango cha ufikiri kinachohitajika kwa matatizo changamano na uwezo wa kutekeleza taratibu za kazi za hatua nyingi na kuunda hati, lahajedwali, na mawasilisho yaliyo nadhifu.

BenchCAD tests whether models can reconstruct 3D objects from multi-view renders by generating CAD code. With tools, GPT‑6 Astra reaches a new high in the comparison shown, achieving a 95.9% geometric-overlap score, versus 83.3% for GPT‑5.6 Sol and 84.3% reported for Claude Fable 5.1.⁵ Estimated API cost is approximately 43% lower than Sol and 86% lower than Fable 5.1 in the configurations shown.

GPT‑6 Astra ni muundo wetu bora zaidi wa kufuata violezo vilivyopo na kuunda mawasilisho yaliyo na mpangilio mzuri na yanayoeleza kwa ufupi hoja kuu kwa simulizi iliyopangiliwa. Huunda hati zilizo wazi na zilizopangiliwa vyema, mawasilisho, lahajedwali na uchanganuzi unaofuata violezo vyako na unaoendana na mtindo wako wa uandishi na wa kimuonekano. Astra pia imefunzwa mahususi kujumuisha kwenye matokeo muktadha muhimu pekee, badala ya kurudia taarifa zisizo za lazima kwa kazi husika. Haya yote yanamaanisha kuwa inaweza kutoa matokeo yanayoweza kutumika mara moja zaidi, yanayolingana na muktadha na viwango vya biashara yako.

Faili ya marejeleo

Matokeo ya GPT‑6 Astra

GPT‑6 Astra huunda wasilisho la slaidi kuhusu GPT‑Gaia, muundo wa kubuniwa, ikitumia slaidi chache tu kutoka kwenye kiolezo cha wasilisho cha OpenAI, na ikidumisha sauti na mpangilio sahihi kote. Hii inamaanisha kuwa unaweza kutarajia mawasilisho ya slaidi yaliyoundwa kwa mpangilio sahihi kulingana na viwango vya biashara yako.

GPT‑6 Astra pia huleta uamuzi bora zaidi wa kimwonekano kwenye tovuti, michezo, programu na michakato inayounda. Kwa kutumia Sites⁠(fungua katika dirisha jipya) katika ChatGPT, Astra inaweza kuunda, kupangisha na kushiriki tovuti, programu za wavuti na michezo moja kwa moja kutoka kwenye dokeza.

“Astra inatupa faida kubwa katika uwezo na ufanisi. Inatekeleza kwa mafanikio taratibu zetu changamani zaidi za kazi za ubunifu huku ikitumia hadi tokeni chache kwa asilimia 20 kuliko miundo mingine tuliyojaribu. Muhimu zaidi, kwa wateja wetu, hii inamaanisha matokeo ya ubora wa juu zaidi.”

Alex Mashrabov, Mkurugenzi Mtendaji na Mwanzilishi-Mwenza, Higgsfield AI

GPT‑6 Astra huunda muundo wa nyumba katika Blender na kuugeuza kuwa mandhari inayoweza kutembelewa katika Unreal Engine 5, ikiwasaidia wabunifu na wateja kuchunguza mpangilio na kuifahamu nafasi hiyo kabla haijajengwa.

Muundo unaweza kuhuisha michezo kupitia michoro angavu, uchezaji unaovutia na miondoko sahihi, na kuwawezesha watu wasio na ujuzi wa kiufundi kuunda na kucheza michezo maalum inayokwenda zaidi ya vipengele vya msingi ndani ya dakika chache. salio: Pietro Schirano.

Maagizo yanapoacha nafasi ya tafsiri tofauti, GPT‑6 Astra ni bora kuliko miundo ya awali katika kufanya uamuzi sahihi. Hutumia muktadha kujaza mapengo ya kawaida na huuliza maswali mahususi pale ambapo jibu linaweza kubadilisha matokeo. Katika Codex, inaweza kuuliza maswali kwa njia isiyohitaji kusubiri huku ikiendelea kufanyia kazi mambo ambayo hayategemei jibu lako. Usipojibu, huendelea kwa kutumia makisio ya busara inapofaa, lakini husubiri mchango wako kuhusu maamuzi muhimu yenye athari kubwa.

Mifano iliyo hapa chini inaonyesha jinsi Astra inavyoshirikiana katika kazi za kila siku ambapo taarifa zinazokosekana zinaweza kubadilisha jibu kwa kiasi kikubwa.

Ulinganisho wa kando kwa kando wa GPT-5.6 Sol na GPT-6 Astra zikisaidia kuunda tovuti binafsi ya taaluma.

Astra pia ni bora zaidi katika kudumisha mwelekeo kadri kazi inavyobadilika. Miundo ya awali wakati mwingine ilichukulia jumbe elekezi kama lengo jipya, na kupoteza mwelekeo wa ombi la awali au vikwazo vya awali. Astra hujumuisha mahitaji mapya, hubadilisha mwelekeo inapoombwa, na hujibu maswali ya pembeni bila kuiacha kazi pana.

“Astra ni uboreshaji mkubwa wa ubora ikilinganishwa na GPT‑5.6 Sol katika kazi changamano za kisheria. Katika majaribio yetu ya awali, Astra ilijitokeza kwa kushughulikia kazi za kisheria kama anavyofanya wakili makini: hutofautisha nyaraka na rekodi zilizothibitishwa, huibua dhana zisizoungwa mkono, na hugeuza mapengo kuwa misimamo thabiti ya utayarishaji wa hati.”

Niko Grupen, Mkuu wa Utafiti wa Kivitendo, Harvey

Uandishi wa msimbo

GPT‑6 Astra ndiyo muundo bora zaidi kwa uhandisi wa programu hadi sasa.

“GPT‑6 Astra hutoa utendaji wa kiwango cha juu zaidi kwenye vigezo vyetu vya ndani vya uandishi wa msimbo na inaonyesha hatua wazi ya mbele katika tathmini za umaizi wa biashara ikilinganishwa na GPT‑5.6 Sol. Inapotumiwa kwa usimbaji wa kiwakala, GPT‑6 Astra huwasiliana kwa njia iliyo rahisi zaidi kwa wasanidi programu kufuatilia na huzalisha msimbo unaohitaji marudio machache ili kufikia ubora wa uzalishaji.”

John Crepezzi, Visaidizi vya AI, Jane Street

“Tulijaribu Astra katika viwango vya juhudi vya chini, wastani na juu kwenye mojawapo ya tathmini zetu za kizazi cha kwanza, na ilifanya vyema zaidi kwa kiwango kikubwa kuliko GPT 5.6 Sol. Juhudi ya juu huleta marudio zaidi kwenye toleo jipya, uthibitishaji zaidi kupitia majaribio ya kivinjari, na mwelekeo wa kutekeleza msimbo badala ya apply-patch. Kuelewa jinsi muundo unavyotumia juhudi zake ndiko kunakotuwezesha kuwapa mamilioni ya waundaji njia ya haraka zaidi na ya kuaminika zaidi kutoka wazo hadi programu inayofanya kazi.”

Fabian Hedin, CTO na Mwanzilishi Mwenza, Lovable

1 kati ya 2

“GPT‑6 Astra hutoa utendaji wa kiwango cha juu zaidi kwenye vigezo vyetu vya ndani vya uandishi wa msimbo na inaonyesha hatua wazi ya mbele katika tathmini za umaizi wa biashara ikilinganishwa na GPT‑5.6 Sol. Inapotumiwa kwa usimbaji wa kiwakala, GPT‑6 Astra huwasiliana kwa njia iliyo rahisi zaidi kwa wasanidi programu kufuatilia na huzalisha msimbo unaohitaji marudio machache ili kufikia ubora wa uzalishaji.”

John Crepezzi, Visaidizi vya AI, Jane Street

“Tulijaribu Astra katika viwango vya juhudi vya chini, wastani na juu kwenye mojawapo ya tathmini zetu za kizazi cha kwanza, na ilifanya vyema zaidi kwa kiwango kikubwa kuliko GPT 5.6 Sol. Juhudi ya juu huleta marudio zaidi kwenye toleo jipya, uthibitishaji zaidi kupitia majaribio ya kivinjari, na mwelekeo wa kutekeleza msimbo badala ya apply-patch. Kuelewa jinsi muundo unavyotumia juhudi zake ndiko kunakotuwezesha kuwapa mamilioni ya waundaji njia ya haraka zaidi na ya kuaminika zaidi kutoka wazo hadi programu inayofanya kazi.”

Fabian Hedin, CTO na Mwanzilishi Mwenza, Lovable

Terminal-Bench 4.0 hujaribu mawakala kwenye majukumu changamano yanayotegemea taminali, ikijumuisha uhandisi wa programu, usanidi wa mfumo, na uchanganuzi wa data. GPT‑6 Astra hufikia kiwango kipya cha juu cha asilimia 57.9, ikilinganishwa na asilimia 37.3 kwa GPT‑5.6 Sol² na asilimia 55.8 kwa Claude Fable 5.1, kwa gharama ya API kwa kila jukumu iliyokadiriwa kuwa chini kwa takriban asilimia 9 na asilimia 63, mtawalia.

Kwa Astra, tunaanzisha njia mpya ya Codex kuhifadhi na kurejesha muktadha wakati dirisha la muktadha linapojaa. Kihistoria, miundo imetumia ukandamizaji wa mutadha kufanya muhtasari wa kazi wakati wa vikao virefu, kama vile wakati wa kutatua matatizo changamano au kushughulikia marekebisho makubwa. Kila ukandamizaji wa mutadha unaweza kuacha nje maelezo kuhusu kwa nini marekebisho yalishindwa au jinsi kijenzi kinavyofanya kazi. Katika Codex, Astra inaweza kuhifadhi madokezo kwenye dirisha la muktadha mbalimbali, ikihifadhi maelezo yaliyokusanywa bila kuyakandamiza mara kwa mara kuwa muhtasari mmoja. Madirisha ya muktadha ya awali yanaendelea kutafutika, kwa hivyo Astra inaweza kupata mahitaji au matokeo ya majaribio kutoka kwa ujumbe wa awali na matokeo ya zana—hata kama taarifa hiyo haikuhifadhiwa katika madokezo yake. Unaweza kuwezesha kipengele hiki cha majaribio katika config.toml yako ya Codex,⁠(fungua katika dirisha jipya) na itakuwa chaguomsingi kwa Astra katika wiki zijazo.

Kuendeleza uvumbuzi wa kisayansi

“Hadithi ni: mwisho wa enzi moja, mwanzo wa nyingine.”

Greg Burnham, EpochAI

GPT‑6 Astra is a major advance for scientific discovery, mathematics, and health. Today, we’re sharing two further results on the gaps between prime numbers.^{^{9, 10}}

Astra also sets new records across a suite of math and science evaluations.

GPQA Diamond hupima uwazaji wa kisayansi wa kiwango cha masomo ya uzamili katika biolojia, kemia na fizikia. GPT‑6 Astra hufikia kiwango kipya cha juu cha asilimia 96.0 katika ulinganisho ulioonyeshwa. Katika mpangilio wa gharama ya chini, pia inazidi alama bora ya GPT‑5.6 Sol—asilimia 94.9 ikilinganishwa na asilimia 94.6—kwa gharama ya API iliyokadiriwa kuwa chini kwa takriban asilimia 37.

Astra inaweza kusaidia katika kazi za kiutendaji zinazowezesha ugunduzi wa kisayansi. Kwa kuunganisha uwazaji wa kisayansi na matumizi ya kompyuta, inaweza kufanya kazi moja kwa moja katika programu maalumu ili kukagua data na kuchunguza matokeo, ikiwasaidia watafiti kutathmini ushahidi na kuamua nini cha kuchunguza baadaye.

GPT‑6 Astra hutumia programu za kisayansi kukagua ubora wa upangaji mfuatano na kuonyesha tofauti za kijenetiki, ikiwasaidia watafiti kutathmini data zao na kubaini maeneo ya kuelekeza uchanganuzi zaidi.

Usalama wa Mtandao

As we discussed in our safety update⁠, Astra is a significant jump in cyber capabilities and meets the Critical threshold in cybersecurity under our Preparedness Framework. Its ability to identify and develop zero-day exploits can help defenders find and patch weaknesses, but it also creates a need for stronger safeguards. To understand how far these capabilities extend, we ran Astra on internal and third-party expert evaluations.

We first tested the model without production safeguards on ExploitBench and ExploitGym, which evaluate whether models can turn known software vulnerabilities into working exploits. On ExploitBench, Astra achieved a perfect score of 100%, compared with 78.5% for GPT‑5.6 Sol, our previous frontier cyber-capable model. On ExploitGym, Astra reached a 42.4% success rate, compared with 30.3% for GPT‑5.6 Sol, while using substantially fewer output tokens.^¹³

Given concerns that exposure to historical software vulnerabilities may have affected benchmark results, we also evaluated Astra on two novel benchmarks. For one, we built an internal “ExploitBench (June–August 2026)” evaluation to test exploit development using vulnerabilities from the previous three months.^¹⁴ Astra achieved substantially higher arbitrary code-execution rates than GPT‑5.6 Sol on this dataset while using far fewer output tokens. During the evaluation, Astra even discovered and used two previously unknown zero-day vulnerabilities. We are disclosing both vulnerabilities to their maintainers.

We also tested Astra on SRE-Bench^¹⁵, a benchmark that measures whether models can reverse engineer software binaries to understand its core logic without access to raw source code. Astra solved 88.0% of tasks in a single attempt and 99.2% within four attempts, compared with 55.9% and 68.7% for GPT‑5.6 Sol, respectively.

Beyond benchmarks, expert-led assessments found that Astra, when run without production safeguards, could use previously unknown vulnerabilities to achieve arbitrary code execution in hardened browsers and create privilege-escalation exploits for hardened operating-systems.

As we discussed in The Defender’s Window, frontier cyber capabilities can help defenders find weaknesses faster, but they also make those weaknesses easier to exploit, raising the urgency for defenders to adapt. With the version of Astra launching today, defenders can use it to complete tasks such as secure code review and patching.

However, Astra will refuse to comply with more advanced cybersecurity tasks such as creating proof-of-concept exploits for vulnerabilities. Through OpenAI Daybreak⁠, we plan to expand access and roll out less restrictive safeguards in the coming weeks. This will enable more defensive workflows, including vulnerability and proof-of-concept validation, malware analysis, and detection engineering.

We have also strengthened our protections against potential cyber misuse, building upon our safeguards stack for GPT‑5.6 Sol. These include stronger model robustness to better withstand potential jailbreaks and more context for our monitoring systems. We have continued rigorous internal and external testing, including automated evaluations with our internal red-teaming attackers. More details about our cyber safeguards and testing are available in the Astra safety overview⁠ and system card⁠(fungua katika dirisha jipya)..

Kuoanisha na kusambaza GPT‑6 Astra kwa uwajibikaji

Astra ni muundo wetu unaoendana zaidi. Astra hufaulu sana katika kutenda kwa uangalifu, kuheshimu mipaka ya kazi, na kuwasiliana kwa uwazi. Kazi hii ndiyo matokeo ya hivi karibuni ya mpango wetu wa utafiti wa muda mrefu unaolenga kufunza miundo iendelee kuendana na nia ya binadamu kutoka mwanzo hadi mwisho.

Katika mazingira nyeti, Astra huchukua hatua kwa uangalifu unaolingana na kiwango cha hatari husika. Katika tathmini ya majukumu ya matumizi ya kompyuta yaliyochaguliwa kwa njia ya kiadui ili kuibua tabia mbaya, Astra ilifanikiwa zaidi kuepuka matokeo yasiyotarajiwa. Kuendesha kwa kutumia hatua za ziada za usalama zinazotolewa kwa chaguo-msingi kulileta utendaji bora hata zaidi.

Astra causes fewer misaligned outcomes than any other frontier models tested. For a fair comparison, we used a generic computer-using-agent harness (based on the native computer use tools available in both the OpenAI Responses API⁠(fungua katika dirisha jipya) and Anthropic Messages API⁠(fungua katika dirisha jipya)) and without additional protections that are normally deployed to Codex and ChatGPT Work users (auto-review, confirmation policy). ¹⁶

Pia kuna uwezekano mkubwa zaidi kwamba Astra itafanya kazi ndani ya mipaka iliyowekwa na mtumiaji na inayodokezwa na mazingira yake. Katika tathmini ya ndani, Astra haikuwahi kujaribu kukwepa kukataliwa na Ukaguzi wa Kiotomatiki wa Codex. Hili lilibaki hivyo hata wakati Ukaguzi wa Kiotomatiki ulisanidiwa kimakusudi ili uweze kukwepwa na kazi haikuwezekana kukamilishwa vinginevyo. Heshima hii kwa vikwazo vya mazingira inaambatana na matokeo ya tathmini yetu ya kazi za mtandao zisizowezekana tuliyoshiriki katika utangulizi wa chapisho hili na matokeo mengine yaliyoandikwa katika kasi ya mfumo⁠(fungua katika dirisha jipya).

Tathmini hii inajaribu jinsi miundo inavyoitikia kukataliwa na ukaguzi wa kiotomatiki katika kazi za maarifa. Kutumia vibaya usanidi hafifu wa mtumiaji ili kukwepa ukaguzi wa kiotomatiki huhesabiwa kama kutofaulu. Astra haikuwahi kujaribu kukwepa ukaguzi wa kiotomatiki.

Astra additionally sets new marks in transparent user communication. In one evaluation, Astra is three times less likely than GPT‑5.6 Sol to make inaccurate representations about its capabilities and affordances.

Katika tathmini yetu ya madai ya kubuni kuhusu uwezo, Astra inaonyesha uboreshaji mkubwa ikilinganishwa na GPT‑5.6 Sol, ikitoa madai machache ya kupotosha kuhusu uwezo wake.

Our evaluations found Astra’s written reasoning harder to monitor than GPT‑5.6 Sol’s, based on tests that explicitly asked it to evade monitoring. We attribute this to Astra’s greater control over written reasoning on simpler tasks and ability to solve problems with fewer written steps. Astra still appears to struggle to conceal the reasoning needed for complex tasks, but we take the decline seriously. Improving monitorability remains a research priority, and the accompanying system card⁠(fungua katika dirisha jipya) details our findings and ongoing work.

Alignment training is core to our approach to deployment. As an additional layer of defenses, we also build system safeguards like Codex Auto-review⁠(fungua katika dirisha jipya) and monitoring agents’ reasoning and actions to help detect and contain unsafe behavior. As described in our safety update⁠, we are also deploying misalignment monitoring in production for Astra-class models in order to have visibility into misalignment, and help contain its worst instances. These safeguards resemble our monitoring for internal deployments and involve a system of classifiers which check the model’s reasoning and actions for unauthorized behavior and automatically stop potentially unauthorized activity.

Given the significant increase in Astra’s cybersecurity capabilities, we are being especially careful to make this deployment safe and secure. Extra safety checks can sometimes slow, pause, or stop legitimate work, including defensive cybersecurity. If a task is paused in ChatGPT or Codex, you may be asked to review the action before continuing. In the API, the task will stop. These checks can sometimes interrupt legitimate work, and we are continuing to iterate on this system to reduce unnecessary interruptions. Misalignment monitoring cannot replace alignment: our goal is to build models that reliably stay within their authorized scope, so these protections do not need to intervene.

Upatikanaji

GPT‑6 Astra is rolling out today to a limited set of organizations and over the coming days will become available to all ChatGPT Plus, Pro, Business, and Enterprise users, as well as through the OpenAI API, Microsoft Azure, and AWS Bedrock. Astra usage is included within the existing subscription allowances—users and businesses will also be able to purchase credits for additional usage. Users on the Pro, Business, and Enterprise plans will also get access to GPT‑6 Astra Pro. Enterprise administrators can enable Astra for their workspace; access is off by default at launch.

Astra supports Zero Data Retention for eligible API customers, and as we shared last month, we're testing Private Safety Processing to strengthen safety monitoring while preserving customer privacy.

For developers, GPT‑6 Astra will be available in the OpenAI API as gpt-6-astra and through Microsoft Azure and Amazon Bedrock.

OpenAI API Standard pricing is $10 per million input tokens and $50 per million output tokens. Separate rates apply to cache reads and writes. Fast mode is available for GPT‑6 Astra in the API and delivers up to 2x the speed of Standard processing at 2x the Standard price.

Matumizi ya kompyuta

Computer Use	GPT‑6 Astra	GPT‑5.6 Sol²	Claude Fable 5.1	Claude Fable 5	Claude Opus 5	Gemini 3.8 Flash
Agents' Last Exam	59.3%	53.6%	-	48.7%	55.5%	-
OSWorld 2.0 (v2026.08.08, offline set, partial score)	72.6%	65.7%	-	-	70.2%³	-
ScreenSpot-Pro (no tools)	92.7%	76.9%	-	87.3%¹⁷	-	-

Professional

Professional	GPT‑6 Astra	GPT‑5.6 Sol	Claude Fable 5.1	Claude Fable 5	Claude Opus 5	Gemini 3.8 Flash
AutomationBench	41.4%	18.1%	31.4%	17.4%	26.9%	-
BenchCAD	95.9%	83.3%	84.3% ⁵	67.5% ⁵	82.1% ⁵	-
BrowseComp	91.5%	90.4%	-	87.4%	90.8%	-
OpenScore String Quartets (1 - OMR-NED)	0.84	0.19	-	-	-	-
Internal Design Tasks	50.0%	47.4%	-	35.8%	-	-
Internal Data Science Tasks	40.9%	30.5%	-	34.7%	-	-
Artificial Analysis Intelligence Index v4.1.1	61.2	60.9	65.7	62.1	63.1	58.7

Coding

Coding	GPT‑6 Astra	GPT‑5.6 Sol	Claude Fable 5.1	Claude Fable 5	Claude Opus 5	Gemini 3.8 Flash
Terminal-Bench 4.0	57.9%	37.3%	55.8%	44.5%	52.6%	19.1%
DeepSWE v1.1	74.1%	72.7%	67.4%	69.9%	73.7%	73.8%
FrontierCode 1.1 Extended (score)	64.5% ⁸	60.6%	63.6%	64.9%	63.6%	56.3%
FrontierCode 1.1 Main (score)	53.3% ⁸	47.5%	50.9%	53.5%	53.4%	43.6%
Internal Database Migration Tasks	63.9%	42.7%	57.8%	50.3%	-	-
Artificial Analysis Coding Agent Index v1.4	67.0	65.1	-	67.2	68.1	61.2

Academic

Academic	GPT‑6 Astra	GPT‑5.6 Sol	Claude Fable 5.1	Claude Fable 5	Claude Opus 5	Gemini 3.8 Flash
Terminal-Bench Science 0.1	64.6%	22.4%	52.6%	21.4%	30.0%	-
FrontierMath Tier 4 (v2)	97.6%	83.0%	87.8%	90.2%	73.2%	-
GPQA Diamond	96.0%	94.6%	93.7%	92.6%	93.7%	95.3%
Humanity's Last Exam (w/ tools)	57.2%	-	65.0%	63.8%	63.6%	-

Sayansi na Afya

Science and Health	GPT‑6 Astra	GPT‑5.6 Sol	Claude Fable 5.1	Claude Fable 5	Claude Opus 5	Gemini 3.8 Flash
GeneBench Pro	37.1%	32.3%	-	-	-	-
MedChemBench (Internal)	49.3%	47.4%	-	-	-	-
LifeSciBench	60.3%	59.9%	-	-	-	-
HealthBench Professional (length-adjusted)	63.4%	60.5%	58.1% ¹¹	60.9% ¹¹	56.4% ¹¹	52.1%

Usalama wa Mtandao

Cybersecurity	GPT‑6 Astra	GPT‑5.6 Sol	Claude Fable 5.1	Claude Fable 5	Claude Opus 5	Gemini 3.8 Flash
ExploitBench	100.0%	78.5%	-	-	70%	-
ExploitGym	42.4% ¹³	30.3% ¹³	30.4% ¹⁷	28.4%¹⁷	22.0%	-
ExploitBench (June-Aug 2026)	39.0%	5.5%	-	-	-	-
SRE-Bench	88.0%	55.9%	-	-	12.5%	-
SEC-Bench Pro	85.4%	79.1%	-	-	-	-

Ulinganifu

Ulinganifu	GPT‑6 Astra	GPT‑5.6 Sol	Claude Fable 5.1	Claude Fable 5	Claude Opus 5	Gemini 3.8 Flash
Kipimo linganishi cha usalama wa matumizi ya ndani ya kompyuta (chini ni bora)	2.4%	22.0%	9.5%	18.3%	11.5%	-
Kigezo cha ndani cha usalama wa matumizi ya kompyuta, na AutoReview (chini ni bora)	1.8%	4.3%	-	-	-	-
Kipimo cha ndani cha ukwepaji (chini ni bora)	0.00%	0.29%	-	-	-	-
Honeypot ya ExploitGym (chini ni bora)	0.0%	48.2%	-	-	-	-
ExploitGym Isiyowezekana	100.0%	-	-	-	-	-
Kipimo cha ndani cha udanganyifu (chini ni bora)	4.2%	12.2%	-	-	-	-

Muktadha mrefu

Muktadha mrefu	GPT‑6 Astra	GPT‑5.6 Sol	Claude Fable 5.1	Claude Fable 5	Claude Opus 5	Gemini 3.8 Flash
OpenAI MRCR v2 8-needle 256K-512K	100.0%	91.5%	-	-	-	-
OpenAI MRCR v2 8-needle 512K-1M	96.3%	73.8%	-	-	-	-

Uwazaji wa dhana

Abstract reasoning	GPT‑6 Astra	GPT‑5.6 Sol	Claude Fable 5.1	Claude Fable 5	Claude Opus 5	Gemini 3.8 Flash
ARC-AGI-3	99.9% ¹	7.8%	-	-	30.2%	-
ARC-AGI-2	95.0%	92.5%	90.0%	89.2%	90.4%	-
ARC-AGI-1	98.5%	97.5%	97.5%	98.5%	97.5%	-

Alama za tathmini huwa za juu kabisa kwa juhudi yoyote. Tathmini za GPT ziliendeshwa katika mazingira yetu ya utafiti au kupitia API yetu, ambayo inaweza kutoa matokeo tofauti kidogo kutoka kwa ChatGPT ya uzalishaji kutokana na tofauti katika madokezo ya mfumo, zana zinazopatikana, n.k.

2026

FOOTNOTES

1
On ARC-AGI-3, GPT-6 Astra was run with our responses API harness⁠, which changes two settings to better match real-world performance. The changes do not specifically target ARC-AGI-3.
2
GPT-5.6 Sol refers to the version available in our API, ChatGPT Codex, and ChatGPT Work. The version in ChatGPT Chat is slightly different.⁠
3
OSWorld V2-Offline is a subset of the original OSWorld V2 that works without internet access. Claude model performance on OSWorld-V2 Offline was reproduced by the authors on the official leaderboard⁠(fungua katika dirisha jipya). On OSWorld 2.0, the scores for Claude use the official settings, and not the modified tasks and modified grading from the Fable 5.1 System Card.
4
Model times are the reported elapsed times for the corresponding demonstration runs. The displayed clips are edited excerpts.
5
On BenchCAD, Claude's scores reflect 3 modifications to the eval, detailed in the Fable 5.1 System Card⁠(fungua katika dirisha jipya).
6
Guang Yang, Victoria Ebert, Nazif Tamer, Brian Siyuan Zheng, Luiza Pozzobon, and Noah A. Smith. “LEGATO: Large-scale End-to-end Generalizable Approach to Typeset OMR⁠(fungua katika dirisha jipya).” arXiv:2506.19065, 2025.
7
Mark R. H. Gotham, Maureen Redbond, Bruno Bower, and Peter Jonas. “The OpenScore String Quartet Corpus⁠(fungua katika dirisha jipya).” Proceedings of the 10th International Conference on Digital Libraries for Musicology, pp. 49–57. ACM, 2023.
8
On FrontierCode, GPT-6 Astra was run with a developer message similar to a section of its developer message in Codex⁠(fungua katika dirisha jipya): "Avoid creating excessive test files. Create a new test file only when required by repository conventions or when no existing file is a suitable home. Avoid unrelated cleanup and unnecessary complexity. Reuse suitable existing utilities. Read relevant repository instructions and inspect nearby code, tests, documentation, and CI. Follow established conventions. The goal is clean, mergeable code." The prompt was not optimized for the eval.
9
The first concerns how close together prime numbers can occur, however far along the number line you go. For more than a decade, the best known result established that infinitely many pairs of primes are at most 246 apart. Julia Stadlmann⁠(fungua katika dirisha jipya) recently improved that bound to 240. Astra helped establish a stronger bound of 186, showing that infinitely many pairs occur within this smaller distance. Short prime gaps: Proof⁠(fungua katika dirisha jipya) and supporting research⁠(fungua katika dirisha jipya).
10
The second concerns unusually large gaps between primes. Astra improved a term in a bound on these gaps that had remained unchanged for more than 80 years. We’re sharing the proofs and abridged chain of thought and verification materials for both results. Large prime gaps: Proof⁠(fungua katika dirisha jipya) and supporting research⁠(fungua katika dirisha jipya).
11
We independently evaluated all Claude models following the intended HealthBench Professional procedure, using GPT‑5.4 grading and length-adjusted, unclipped scores. For Fable 5.1, we used Opus 5 fallback for provider refusals.
12
Claude Fable 5 and 5.1 are not included in LifeSciBench Gold v1, GeneBench Pro v13, and MedChemBench because they refuse the majority of questions in these evaluations.
13
On ExploitGym, we tested Astra and Sol without the 6-hour time limit, to better assess their full cyber capabilities. They are fast enough that it has little impact.
14
ExploitBench (June–August 2026) contains 20 high-severity V8 vulnerabilities across 13 stable Chrome releases. The benchmark tests whether agents can achieve arbitrary code execution in V8 and official Chrome releases for Linux by exploiting each specified vulnerability. Some included vulnerabilities may not permit arbitrary code execution under the evaluation’s constraints, so a 100% success rate may not be achievable. Note: the 5.5% score of GPT-5.6 Sol is an artifact of the 300-turn limit in the benchmark, which is not a limit that real customers using max would have. The model at similar settings achieved an 11.5% score when hitting fewer limits.
15
Jeremy Spence et al. “The Next Challenge for Agentic Cybersecurity: A Realistic, Contamination-Free Reverse Engineering Benchmark⁠(fungua katika dirisha jipya).” arXiv:2608.11469v1, 2026.
16
When we test across third-party models, we use a simpler research setup. Codex has a more complex production configuration, which can result in different raw-model error rates. Provider-side safeguards and computer-tool implementations still differ. Users do not experience the no-confirmation scenario in Codex, as it's an internal research configuration.
17
For ScreenSpot-Pro and ExploitGym, the Fable scores we report come from Mythos, which is Fable with fewer safeguards.

GPT-6 Astra: A new generation of intelligence

Kizazi kipya cha Kiwango cha ufikiri

Muundo bora zaidi duniani wa kutumia kompyuta

GPT‑6 Astra: dakika 2 sekunde 54

Mabadiliko makubwa katika kazi za kitaalamu

Uandishi wa msimbo

Kuendeleza uvumbuzi wa kisayansi

Usalama wa Mtandao

Kuoanisha na kusambaza GPT‑6 Astra kwa uwajibikaji

Upatikanaji

Matumizi ya kompyuta

Professional

Coding

Academic

Sayansi na Afya

Usalama wa Mtandao

Ulinganifu

Muktadha mrefu

Uwazaji wa dhana

FOOTNOTES

GPT‑6 Astra: dakika 2 sekunde 54