GPT-6 Astra: A new generation of intelligence
Nova generacija inteligencije
Predstavljamo GPT‑6 Astra, najinteligentniji i najusklađeniji model na svijetu.
GPT‑6 Astra objedinjuje godine istraživanja i velika ulaganja u predobuku, učenje sa potkrepljivanjem i usklađivanje. Astra postiže najsavremenije rezultate u korištenju računara, pregledanju interneta, softverskom inženjerstvu, kibernetičkoj sigurnosti, nauci i profesionalnom radu. Astra dostiže zasićenje na FrontierMath nivou 4 rezultatom od 98%, nakon što je već pomogla u rješavanju dugogodišnjih otvorenih problema u matematici. Astra također postiže rezultat od 99,9% na ARC-AGI-3 i rezultat od 100% na ExploitBenchu. Također postavlja novu granicu u korištenju računara i preglednika, obavljajući najzahtjevniji profesionalni rad nenadmašnom brzinom, tačnošću i prosuđivanjem.
GPT‑6 Astra se danas uvodi ograničenom broju organizacija, a tokom narednih dana postaće dostupan svim korisnicima ChatGPT Plus, Pro, Business i Enterprise, kao i putem OpenAI API-ja, Microsoft Azurea i AWS Bedrocka.
Astra je naš najusklađeniji model, sa značajnim poboljšanjima u razumijevanju namjere korisnika i ponašanja modela — zadatke možete delegirati sa većim povjerenjem u Astrino prosuđivanje. Kao jedan od načina na koji ovo testiramo, razvili smo novu procjenu zasnovanu na incidentu s platformom Hugging Face, koja provjerava hoće li model suočen sa teškim ili nemogućim zadatkom prekoračiti svoj predviđeni djelokrug. U poređenju s modelom GPT‑5.6 Sol, koji je bez zaštitnih mjera u produkciji u 48% slučajeva prekoračio odobreni cilj, GPT‑6 Astra je to učinio u 0% slučajeva.
Najbolji svjetski model za korištenje računara
GPT‑6 Astra označava novu granicu u brzini, tačnosti i sigurnosti korištenja računara. Može se pobrinuti za zamorne zadatke kao što su popunjavanje online obrazaca, ažuriranje podataka o klijentima u CRM-u i organiziranje vašeg kalendara. Može provoditi istraživanja na internetu i izrađivati nacrte sažetaka u vašoj e-pošti ili u uređivaču dokumenata. Može analizirati naučne podatke, generirati grafikone, kreirati web-stranicu i pokretati frontend QA provjere kako bi se osiguralo da sve funkcije na toj stranici rade. Može vam pomoći da autonomno instalirate i testirate softver te otklanjate probleme koje vidite na ekranu. Ova poboljšanja se odražavaju i u našim najsavremenijim rezultatima evaluacije.
Ova poboljšanja također dovode do značajnog povećanja efikasnosti u stvarnim zadacima rada na temelju znanja. U simulacijama kašnjenja na OSWorld 2.0, Astra ostvaruje bolje performanse pri korištenju računara uz oko 47% manje vremena po zadatku od GPT‑5.6 Sol, sa rezultatom od 72,6% za otprilike 40 minuta po zadatku, u poređenju sa 65,7% za otprilike 75 minuta.3
Mogućnosti GPT‑6 Astre za korištenje računara mogu se vidjeti u rezultatima u različitim oblastima, uključujući razvoj igara, elektrotehniku i svakodnevni rad zasnovan na znanju:
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
Značajan iskorak u profesionalnom radu
GPT‑6 Astra kombinuje napredak u korištenju računara sa ciljanom obukom za profesionalna okruženja kako bi pomogao u rješavanju složenih radnih zadataka. Kombinuje inteligenciju potrebnu za rješavanje složenih problema s mogućnošću izvršavanja radnih tokova u više koraka i izrade dotjeranih dokumenata, proračunskih tabela i prezentacija.
GPT‑6 Astra je naš najbolji model za pridržavanje postojećih predložaka i izradu dobro raspoređenih slajdova koji sažeto prenose ključne tačke kroz strukturiranu naraciju. Kreira jasne, dobro strukturirane dokumente, prezentacije, proračunske tabele i analize koje prate vaše predloške i odgovaraju vašem stilu pisanja i vizuelnom stilu. Astra je također obučena da u izlazne rezultate posebno uključuje samo relevantni kontekst, umjesto da ponavlja informacije koje nisu potrebne za trenutni zadatak. Sve ovo znači da može generisati više odmah upotrebljivih artefakata koji odgovaraju vašem poslovnom kontekstu i standardima.
GPT‑6 Astra također unosi snažniji vizuelni sud u web-stranice, igre, aplikacije i rendere koje izrađuje. Uz Sites(otvara se u novom prozoru) u ChatGPT‑u, Astra može kreirati, hostovati i dijeliti web-stranice, web-aplikacije i igre direktno iz upita.
Kada uputstva ostavljaju prostor za tumačenje, GPT‑6 Astra bolje od prethodnih modela donosi ispravnu odluku. Koristi kontekst za popunjavanje uobičajenih praznina i postavlja ciljana pitanja kada bi odgovor mogao promijeniti ishod. U Codexu može postavljati pitanja asinhrono dok nastavlja s radom koji ne zavisi od vašeg odgovora. Ako ne odgovorite, nastavlja na osnovu razumnih pretpostavki gdje je to prikladno, ali čeka vaš unos za odluke s važnim posljedicama.
Primjeri u nastavku prikazuju kako Astra sarađuje na svakodnevnim zadacima kod kojih informacije koje nedostaju mogu značajno promijeniti odgovor.
Astra je također bolja u održavanju orijentacije kako se zadatak razvija. Raniji modeli su poruke za usmjeravanje ponekad tretirali kao novi cilj, gubeći iz vida prvobitni zahtjev ili ranija ograničenja. Astra uključuje nove zahtjeve, mijenja smjer kada se to zatraži i odgovara na usputna pitanja bez zanemarivanja šireg zadatka.
Kodiranje
GPT‑6 Astra je dosad najbolji model za softversko inženjerstvo.
“GPT‑6 Astra pruža vrhunske performanse na našim internim referentnim testovima kodiranja i pokazuje jasan napredak u evaluacijama intuicije za trgovanje u poređenju s GPT‑5.6 Sol. Kada se koristi za kodiranje s agentima, GPT‑6 Astra komunicira na način koji programeri lakše prate i proizvodi kod koji zahtijeva manje iteracija da bi dostigao kvalitet za produkciju.”
„Testirali smo Astru pri niskom, Srednja i Visoka nivou napora na jednoj od naših evaluacija prve generacije i značajno je nadmašila GPT 5.6 Sol. Veći nivo napora donosi više iteracija na novoj verziji, više provjere putem testiranja u pretraživaču i veću sklonost izvršavanju koda umjesto primjeni zakrpa. Razumijevanje načina na koji model ulaže svoj napor omogućava nam da milionima graditelja pružimo brži i pouzdaniji put od ideje do funkcionalne aplikacije.”
S Astrom uvodimo novi način da Codex sačuva i preuzme kontekst kada se kontekstni prozor popuni. Historijski gledano, modeli su koristili kompakciju za sažimanje rada tokom dugih sesija, kao što su otklanjanje složenih problema ili rad na velikim refaktorisanjima. Svaka kompakcija može izostaviti detalje o tome zašto ispravka nije uspjela ili kako se komponenta ponaša. U Codex Astra može čuvati bilješke kroz više kontekstnih prozora, čuvajući prikupljene detalje bez njihovog ponovljenog sažimanja u jedan sažetak. Raniji kontekstni prozori ostaju pretraživi, tako da Astra može pronaći zahtjeve ili rezultate testiranja iz prethodnih poruka i izlaza alata—čak i ako te informacije nisu zabilježene u njenim bilješkama. Ovu eksperimentalnu funkciju možete omogućiti u svojoj datoteci Codex config.toml,(otvara se u novom prozoru) i postat će zadana opcija za Astru u narednim sedmicama.
Unapređivanje naučnih otkrića
Astra može pomoći u praktičnom radu koji stoji iza naučnih otkrića. Kombinovanjem naučnog rezonovanja s korištenjem računara može direktno raditi u specijaliziranom softveru kako bi pregledala podatke i istraživala rezultate, pomažući istraživačima da procijene dokaze i odluče šta sljedeće istražiti.
Kibernetička sigurnost
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-Bench15, 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(otvara se u novom prozoru)..
Odgovorno usklađivanje i primjena GPT‑6 Astra
Astra je naš najusklađeniji model. Astra se ističe u pažljivom postupanju, poštivanju granica zadatka i transparentnoj komunikaciji. Ovaj rad najnoviji je rezultat našeg dugogodišnjeg istraživačkog programa usmjerenog na obuku modela koji ostaju usklađeni sa ljudskom namjerom od početka do kraja.
U osjetljivim okruženjima, Astra postupa s oprezom srazmjernim riziku. U evaluaciji zadataka korištenja računara koji su adverzarijalno odabrani kako bi izazvali nepoželjno ponašanje, Astra je bila uspješnija u izbjegavanju nenamjeravanih posljedica. Pokretanje uz dodatne sigurnosne mjere koje se nude prema zadanim postavkama rezultiralo je još boljim performansama.
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(otvara se u novom prozoru) and Anthropic Messages API(otvara se u novom prozoru)) and without additional protections that are normally deployed to Codex and ChatGPT Work users (auto-review, confirmation policy). 16
Također je vjerovatnije da će Astra djelovati unutar granica koje je postavio korisnik i koje podrazumijeva njeno okruženje. U internoj procjeni Astra nikada nije pokušala zaobići odbijanje automatskog pregleda Codexa. Ovo je i dalje važilo čak i kada je Auto-review namjerno podešen tako da se mogao zaobići, a zadatak inače nije bilo moguće dovršiti. Ovo poštovanje ograničenja okruženja u skladu je s rezultatima naše evaluacije nemogućih kibernetičkih zadataka, koje smo podijelili u uvodu ovog teksta, i drugim nalazima dokumentovanim u našoj kartici sistema(otvara se u novom prozoru).
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.
U našoj evaluaciji halucinacija o sposobnostima, Astra pokazuje znatno poboljšanje u odnosu na GPT‑5.6 Sol, iznoseći manje obmanjujućih tvrdnji o svojim sposobnostima.
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(otvara se u novom prozoru) 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(otvara se u novom prozoru) 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.
Dostupnost
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.
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 |
Akademski
Akademski | GPT‑6 Astra | GPT‑5.6 Sol | Claude Fable 5.1 | Claude Fable 5 | Claude Opus 5 | Gemini 3.7 Flash |
GPQA Diamond | 96,0% | 94,6% | 93,7% | 92,6% | 93,2% | 94,5% |
Teorijsko računarstvo | 75,4% | |||||
FrontierMath nivo 1-3 (v2) | 89,0% | 90,2% | 87,0% | 85,6% | 71,6% | |
FrontierMath nivo 4 (v2) | 97,6% | 83,0% | 87,8% | 87,8% | 73,2% | 36,6% |
Posljednji ispit čovječanstva (alati) | 65,0% | 63,8% | 63,6% | |||
Posljednji ispit čovječanstva (bez alata) | 59,1% | 55,5% | 54,9% | 47,9% |
Usklađenost
Usklađenost | GPT‑6 Astra | GPT‑5.6 Sol | Claude Fable 5.1 | Claude Fable 5 | Claude Opus 5 | Gemini 3.8 Flash |
Sigurnosni referentni pokazatelj za internu upotrebu računara (niže je bolje) | 2,4% | 22,0% | 9,5% | 18,3% | 11,5% | - |
Interni referentni test sigurnosti pri korištenju računara, uz AutoReview (niže je bolje) | 1,8% | 4,3% | - | - | - | - |
Interna referentna vrijednost zaobilaženja (niža je bolja) | 0,00% | 0,29% | - | - | - | - |
Honeypot u ExploitGymu (niže je bolje) | 0,0% | 48,2% | - | - | - | - |
Nemogući ExploitGym | 100,0% | - | - | - | - | - |
Interni benchmark halucinacija (niži je bolji) | 4,2% | 12,2% | - | - | - | - |
Dugi kontekst
Dugi kontekst | 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-igala 256K-512K | 100,0% | 91,5% | - | - | - | - |
OpenAI MRCR v2 8-needle 512K-1M | 96,3% | 73,8% | - | - | - | - |
Apstraktno rezonovanje
| 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% | - |
Ocjene evaluacije su maksimalne pri bilo kojem nivou napora. Evaluacije GPT‑a su provedene u našem istraživačkom okruženju ili putem našeg API-ja, što može dati nešto drugačiji rezultat od produkcijskog ChatGPT‑a zbog razlika u sistemskim upit, dostupnim alatima itd.
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(otvara se u novom prozoru). 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
- 5
On BenchCAD, Claude's scores reflect 3 modifications to the eval, detailed in the Fable 5.1 System Card(otvara se u novom prozoru).
- 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(otvara se u novom prozoru).” arXiv:2506.19065, 2025.
- 7
Mark R. H. Gotham, Maureen Redbond, Bruno Bower, and Peter Jonas. “The OpenScore String Quartet Corpus(otvara se u novom prozoru).” 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(otvara se u novom prozoru): "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(otvara se u novom prozoru) 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(otvara se u novom prozoru) and supporting research(otvara se u novom prozoru).
- 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(otvara se u novom prozoru) and supporting research(otvara se u novom prozoru).
- 11
- 12
- 13
- 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(otvara se u novom prozoru).” 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
