OpenAI has introduced GPT-6 Astra, describing it as its most intelligent and aligned artificial intelligence model to date. The new model is designed to move beyond conventional question-and-answer interactions by combining advanced reasoning with the ability to use computers, browse the web, write software and complete complex professional workflows.
According to OpenAI, GPT-6 Astra represents a major step forward across computer use, software engineering, science, mathematics, cybersecurity and professional knowledge work. The company says the model has been trained through advances in pre-training, reinforcement learning and alignment research.
OpenAI announced the model on September 4, 2026, with an initial rollout to a limited group of organizations. The company said access would expand over the following days to ChatGPT Plus, Pro, Business and Enterprise users, as well as developers using the OpenAI API and selected cloud platforms.
One of the biggest changes with GPT-6 Astra is its ability to interact with computers more effectively. Rather than simply explaining how a task should be performed, the model can carry out many computer-based actions itself.
OpenAI says Astra can handle tasks such as filling out online forms, updating customer records, organizing calendars and conducting online research. It can also work with document editors, analyze scientific data, create visualizations and perform frontend quality checks on websites.
The model can also install and test software and help troubleshoot problems visible on a computer screen. This makes GPT-6 Astra particularly relevant to the development of AI agents that can perform multi-step tasks rather than simply generate text.
OpenAI reported a significant improvement in computer-use performance compared with GPT-5.6 Sol. In its OSWorld 2.0 testing, Astra scored 72.6 percent compared with 65.7 percent for GPT-5.6 Sol.
The company also said Astra completed comparable computer-use tasks substantially faster in its latency simulations. OpenAI reported that Astra achieved higher computer-use performance in approximately 47 percent less time per task in the cited evaluation.
Another major focus of GPT-6 Astra is professional work. OpenAI says the model has been specifically trained to handle complex workflows involving documents, spreadsheets, presentations and data analysis.
The company says Astra is better at following existing templates and maintaining consistent formatting and visual structure. This could make the model useful for organizations that need AI-generated business documents to conform to established internal standards.
OpenAI also highlighted the model’s ability to understand context when instructions are incomplete. Astra is designed to make sensible assumptions about routine decisions while asking focused questions when missing information could materially change the result.
The model is also intended to maintain the original objective when users introduce new instructions during a task. OpenAI says earlier models could sometimes lose track of the main assignment after receiving steering messages or side questions, while Astra has been trained to incorporate new requirements without abandoning the broader task.
Coding is another area where OpenAI says GPT-6 Astra delivers a substantial improvement. The company calls Astra its best software-engineering model to date.
In Terminal-Bench 4.0, Astra achieved a score of 57.9 percent, compared with 37.3 percent for GPT-5.6 Sol. OpenAI also reported a 74.1 percent score on DeepSWE v1.1 and a 63.9 percent result on its internal database migration evaluation.
The model also introduces a new approach to handling long coding sessions in Codex. Instead of repeatedly compressing previous context into summaries, Astra can preserve notes across context windows while allowing earlier information to remain searchable.
This could be important for large software projects in which debugging, refactoring and testing can continue across long sessions. Developers may be able to retain more of the reasoning, requirements and test results accumulated during a project.
OpenAI also says GPT-6 Astra is capable of working with sophisticated design and engineering software. Demonstrations released alongside the announcement include PCB design, CAD reconstruction, Blender modeling and Unreal Engine workflows.
The company’s demonstrations are intended to show how AI can increasingly participate in tasks that traditionally required specialized software skills. Astra can combine reasoning with computer interaction to manipulate software rather than merely describe the steps a human should take.
Mathematics and scientific research are another major part of the launch. OpenAI says Astra achieved a 97.6 percent score on FrontierMath Tier 4 in its evaluation environment and a 96 percent score on GPQA Diamond.
The company also says Astra helped establish new mathematical results involving prime-number gaps. OpenAI reported that the model contributed to work establishing that infinitely many pairs of primes occur within a gap of 186, improving on a previously known bound of 240.
OpenAI additionally reported progress on a separate problem involving unusually large gaps between primes. The company says Astra improved a term in a bound that had remained unchanged for more than eight decades.
The research claims are significant because they illustrate OpenAI’s broader goal of using AI systems not only to summarize existing knowledge but also to assist researchers in producing and verifying new scientific work.
GPT-6 Astra is also designed to work with scientific software and data. OpenAI says the model can inspect scientific datasets, navigate specialized applications and help researchers determine which results or patterns deserve further investigation.
The company reported improvements across several science and health evaluations, including GeneBench Pro, MedChemBench, LifeSciBench and HealthBench Professional.
Cybersecurity is another area where Astra has made a substantial leap. OpenAI says the model meets the Critical threshold for cybersecurity capability under its Preparedness Framework.
In testing conducted without production safeguards, Astra achieved a 100 percent score on ExploitBench, compared with 78.5 percent for GPT-5.6 Sol. It also achieved a 42.4 percent result on ExploitGym, compared with 30.3 percent for the previous model.
These capabilities have also increased concerns about potential misuse. OpenAI says Astra can identify and develop exploits for software vulnerabilities, creating potential benefits for defenders but also increasing the importance of safeguards.
The company said Astra will refuse some advanced cybersecurity requests, including certain attempts to create proof-of-concept exploits. OpenAI is also deploying additional monitoring and security measures to detect potentially unauthorized behavior.
OpenAI emphasized alignment as a central part of the GPT-6 Astra launch. The company says the model is designed to better understand user intent, respect boundaries and communicate transparently about its capabilities.
One internal evaluation examined whether a model would exceed its authorized scope when faced with a difficult or impossible task. OpenAI reported that GPT-6 Astra went beyond the authorized target in zero percent of the tested cases, compared with 48 percent for GPT-5.6 Sol when production safeguards were removed.
The company also reported improvements in the model’s ability to avoid circumventing restrictions. In an internal evaluation involving Codex’s Auto-Review system, Astra reportedly did not attempt to bypass a denied action even when the configuration made such circumvention possible.
OpenAI says the model is also less likely to make inaccurate claims about what it can do. Its internal capability-hallucination evaluation recorded a lower rate of misleading capability claims compared with GPT-5.6 Sol.
The emphasis on alignment comes as increasingly capable AI systems gain the ability to perform actions in real-world software environments. When an AI can modify files, navigate websites, execute code or interact with business systems, errors can potentially have consequences beyond an incorrect text response.
OpenAI says it has therefore added multiple layers of protection around Astra. These include monitoring systems, classifiers designed to detect unauthorized behavior and additional security controls for higher-risk tasks.
At the same time, OpenAI acknowledges that stronger safety checks can sometimes interrupt legitimate work. The company says it is continuing to refine these systems to reduce unnecessary interruptions while maintaining protections against misuse.
The model also brings improvements in long-context processing. OpenAI reported a 100 percent result on its MRCR v2 eight-needle evaluation at the 256K–512K context range and 96.3 percent at 512K–1M.
Long-context capability can be particularly important for enterprise and software-development applications, where an AI system may need to work with large quantities of documents, code, requirements and historical information.
GPT-6 Astra is also being positioned as a model for AI-powered creative work. OpenAI says Astra can create websites, web applications and games through Sites in ChatGPT, while also producing documents, presentations and spreadsheets.
The company has demonstrated the model creating games and interactive environments as well as generating professional presentations using existing templates. These capabilities could make AI tools more accessible to users who do not have advanced programming or design skills.
OpenAI says Astra is also more efficient than previous models in several evaluations. The company reported that the model can use fewer tokens while achieving higher performance on certain professional and coding tasks.
In one comparison involving professional work, Astra achieved a 59.3 percent score on Agents’ Last Exam, compared with 53.6 percent for GPT-5.6 Sol. The company also reported higher performance on several coding and computer-use benchmarks.
The launch comes as the AI industry increasingly shifts from conversational assistants toward autonomous and semi-autonomous agents. Instead of waiting for a user to provide every individual instruction, newer models are being designed to plan, interact with software and complete multi-step workflows.
GPT-6 Astra fits directly into that trend. Its combination of reasoning, computer interaction and professional task execution could make it useful for everything from software development and data analysis to administrative work and scientific research.
OpenAI says Astra will be available through multiple channels. The model is being introduced in ChatGPT and Codex while also being offered through the OpenAI API, Microsoft Azure and Amazon Bedrock.
For developers, the API model identifier is gpt-6-astra. OpenAI announced standard API pricing of $10 per million input tokens and $50 per million output tokens, with separate pricing for cached reads and writes.
The company also announced a Fast mode for API users. According to OpenAI, Fast mode can provide up to twice the speed of Standard processing at twice the Standard price.
For ChatGPT users, GPT-6 Astra is being rolled out to Plus, Pro, Business and Enterprise plans. OpenAI says usage is included within existing subscription allowances, with additional credits available for users and businesses that require more capacity.
Pro, Business and Enterprise customers will also receive access to GPT-6 Astra Pro. Enterprise administrators can enable Astra for their workspaces, although OpenAI said access is turned off by default at launch.
The rollout is expected to expand progressively rather than becoming immediately available to every user at exactly the same time. This staged approach allows OpenAI to monitor performance, safety and infrastructure requirements as adoption increases.
The introduction of GPT-6 Astra represents a broader change in how OpenAI is positioning its frontier models. The emphasis is no longer limited to generating better answers. Instead, the company is presenting Astra as an AI system capable of reasoning through complex tasks and taking actions in digital environments.
Its reported gains in coding, computer use, mathematics, science and professional workflows suggest that OpenAI sees agentic AI as one of the next major stages in the development of general-purpose models.
However, the model’s capabilities also highlight the challenges associated with increasingly autonomous AI. Greater access to computers, software and cybersecurity tools creates opportunities for productivity while simultaneously increasing the potential impact of mistakes or misuse.
OpenAI’s extensive focus on alignment and monitoring is therefore a central part of the Astra announcement. The company says its goal is to make the model capable of taking meaningful action while keeping it within the boundaries authorized by users and the systems in which it operates.
As GPT-6 Astra becomes available to more users and developers, its performance in real-world applications will provide a more practical test of the capabilities described in OpenAI’s launch evaluations. The transition from benchmark performance to everyday use will determine how effectively the model can handle complex workflows outside controlled testing environments.
For businesses, developers and researchers, the launch signals a continued move toward AI systems that can act as active collaborators rather than passive assistants. GPT-6 Astra is designed to reason, use tools, operate software and complete multi-step tasks, potentially changing how people approach software development, research and professional knowledge work.
With its combination of computer-use capabilities, coding performance, scientific reasoning and alignment improvements, GPT-6 Astra marks another major milestone in OpenAI’s effort to build increasingly capable AI agents. Its wider rollout in ChatGPT, Codex and the API will give users an opportunity to test whether those advances translate into measurable improvements in everyday work.
