OpenAI unveiled GPT-6 Astra on September 3, 2026, describing it as the company’s most powerful model to date. Although the name “ChatGPT Astra” also appears online, GPT-6 Astra is more accurate: it is a new model that OpenAI is gradually making available to ChatGPT users while also integrating it into its tools for work, programming and application development.
GPT-6 Astra delivers its biggest advances in computer control, web use, programming and longer autonomous tasks. Rather than merely responding in a chat, it can work with applications, create documents, presentations and websites, and continue handling a complex task even when the instructions change. Availability is initially limited, with a gradual rollout planned for paid ChatGPT plans.
What is GPT-6 Astra?
GPT-6 Astra is a new generation of OpenAI’s large language model and the successor to the GPT-5.6 family. OpenAI describes it as both its most intelligent and best-aligned model yet. Alignment refers to the model’s ability to correctly understand the user’s intent, stay within the boundaries of the instructions and avoid taking actions the user did not request.
You might also like
A key change is that Astra is no longer presented merely as a better chatbot. OpenAI is placing a strong emphasis on what is known as agentic work. Instead of simply explaining how something could be done, the model can carry out the individual steps when given appropriate access to tools.
For example, it can work with a browser, documents, programming tools and other applications, gradually completing a task involving dozens of steps.
The biggest new feature: Astra can operate a computer
The most significant improvement concerns computer use—the AI’s ability to operate a computer’s graphical interface in much the same way as a person.
OpenAI’s examples include filling in forms, updating CRM records, organizing a calendar, conducting online research, working with documents and checking completed websites.
According to OpenAI, Astra can, for example:
- browse the web and gather information,
- work with forms,
- edit data in applications,
- create and review websites,
- use software through its graphical interface,
- install and test programs,
- respond to a problem it sees on the screen,
- work with documents, spreadsheets and presentations,
- carry out multi-step workflows.
This is not merely a laboratory demonstration. OpenAI says Astra scored 72.6% on OSWorld 2.0, compared with 65.7% for GPT-5.6 Sol. The time difference is particularly notable: Astra completed the simulated task in approximately 40 minutes, while the previous model took around 75 minutes.
That matters because, with AI agents, the question is not only whether they can complete a task, but also how much time and how many steps they require.
It is no longer just one prompt and one answer
When using older models, it was relatively easy to knock them off course. A user could assign a complex task and then add a new requirement while the model was working, causing it to sometimes ignore the original instructions or focus only on the latest message.
GPT-6 Astra is expected to be significantly better in this respect. It can adjust its approach as it works without losing sight of the original goal and earlier constraints.
For example, if you ask the AI to create a website, then tell it to simplify the design, later change the structure of one section and finally ask it to check the mobile version, Astra is expected to incorporate all these requirements into one ongoing task.
OpenAI calls this capability mid-turn steering. For users, however, the simpler explanation is more useful: the model can be corrected and guided while it is working.
It can work on several things at once
Another new capability is the ability to continue with one part of a task while waiting for other information or the result of a tool.
If the model needs additional input from the user, the entire process does not necessarily have to stop. In environments that support this feature, it can continue working on tasks that do not depend on the answer.
Similarly, it can continue reasoning or carry out another part of the assignment while waiting for an external tool.
In practice, this brings AI closer to the way people handle multiple tasks. If we are waiting for a file to download, for example, we do not have to sit and stare at the screen—we can complete another part of the work in the meantime.
Documents, Excel spreadsheets and presentations should be significantly better
With GPT-6 Astra, OpenAI has also placed a strong emphasis on professional office work. The model was trained to create documents, spreadsheets, analyses and presentations that more closely follow an existing template or visual style.
This addresses one of the practical problems with generative AI. Even older models could create presentations, but the result often looked like an entirely new document rather than a natural continuation of an existing corporate presentation.
Astra is expected to better understand the structure of a template, the amount of text, the placement of elements and the context of a document. OpenAI also says it aims to transfer only relevant information into the output rather than mechanically repeating all the context provided.
For companies, the ability to “edit this presentation using our existing template” may therefore be more valuable than a simple request to “create a presentation.”
A major leap in programming
According to OpenAI, GPT-6 Astra is also its best model yet for software engineering.
The model can do more than generate individual pieces of code: it can work with a larger project, run code, test the result, find bugs and then make further changes.
Astra scored 57.9% on the Terminal-Bench 4.0 benchmark, compared with 37.3% for GPT-5.6 Sol. The difference was smaller on DeepSWE v1.1, at 74.1% versus 72.7%.
The results also show why a new AI model should not be judged by a single number. The difference is dramatic in some tests, only a few percentage points in others, and Astra does not have a clear lead over its competitors on some independent measures.
Astra can retain context more effectively during very long tasks
The model itself supports a context window of approximately 1.05 million tokens and a maximum output of up to 128,000 tokens.
A large context window allows the model to work with an exceptionally extensive volume of text, code or documentation at once. However, a practical problem arises with tasks that last even longer.
OpenAI has therefore introduced a mechanism in Codex that allows Astra to create notes and later search older parts of its working context. Older models often gradually compressed a conversation into a brief summary, potentially losing important details in the process.
During a programming session lasting several hours, the difference can be crucial. Information about why a particular attempt failed two hours earlier may later be just as important as the current section of code.
Astra has also made a major leap in science and mathematics
OpenAI published several very high scores from mathematical, scientific and abstract reasoning benchmarks.
Astra scored 97.6% on FrontierMath Tier 4 v2, compared with 83% for GPT-5.6 Sol. It achieved 96% on GPQA Diamond.
The most striking figure comes from ARC-AGI-3. According to OpenAI, Astra scored 99.9%, while GPT-5.6 Sol achieved only 7.8%.
Such a difference looks extremely impressive, but benchmark results should be interpreted with caution. OpenAI itself notes that these results were achieved in a specific testing environment with tools available, and that the production version of ChatGPT may use different system settings.
The benchmark therefore does not mean that the model will be one hundred times better than its predecessor at every real-world problem.
Is GPT-6 Astra already AGI?
The model’s unveiling reignited debate about AGI, or artificial general intelligence. OpenAI President Greg Brockman acknowledged that the period surrounding Astra might one day be viewed as the beginning of the AGI era.
However, this should not be mistaken for scientifically confirmed evidence that OpenAI has created general intelligence at a human level.
Moreover, there is no single universally accepted threshold for AGI. GPT-6 Astra still operates as an AI model with defined tools, limitations and environments. Exceptional benchmark results alone are therefore not enough to declare conclusively that AGI has been achieved.
A much more concrete change is Astra’s ability to work independently for longer periods in real computer applications.
New capabilities also create new security concerns
Cybersecurity is an interesting but also sensitive part of the announcement. GPT-6 Astra is OpenAI’s first model to reach the “Critical” level for cyber capabilities under the company’s Preparedness Framework.
In tests, the model was significantly better at analyzing software vulnerabilities. OpenAI has therefore introduced additional safeguards and restricted some advanced capabilities.
For ordinary users, this does not mean that ChatGPT has gained a mode for carrying out cyberattacks. On the contrary, the production version of Astra has restrictions designed to prevent the misuse of high-risk capabilities. OpenAI says that, in some cases, a legitimate security task may be paused for additional checks, requiring the user to confirm how to proceed.
The company also says Astra is substantially more resistant to prompt-injection attacks, in which content on a website or in a document attempts to manipulate an AI agent into taking an action other than the one requested by the user.
Astra should be less likely to make false claims about what it can do
One of the more interesting safety metrics is not simply whether the model refuses dangerous requests, but whether it accurately reports its own capabilities.
OpenAI says that, in one internal test, Astra was approximately three times less likely than GPT-5.6 Sol to make inaccurate claims about what it could do or what it had just done.
That may sound like a minor detail, but it is extremely important for agentic models. A chatbot that incorrectly says, “the order has been submitted,” is inconvenient. An agent with genuine access to websites, files and applications that falsely claims to have completed an operation could cause a much more serious problem.
What can Astra do in real-world use?
OpenAI’s official presentation shows the direction in which the company wants to take ChatGPT. Instead of a user entering isolated questions into a text box, we see an assistant working across different applications.
In the demonstrations, Astra creates a presentation, prepares an online advertisement, works with a photo stored on the computer, creates a simple game, edits a document, searches for an available appointment and continues with other tasks at the same time.
The most interesting point, therefore, is not that AI can create a presentation or program a game. Previous models could do that as well. The difference is supposed to be its ability to combine these activities into a single workflow and work directly with the user’s tools.
Official GPT-6 Astra demonstration
In its brief introduction, OpenAI shows Astra handling several parallel tasks, including working with a browser, documents, a 3D model and other applications.
When will GPT-6 Astra be available in ChatGPT?
OpenAI began a limited rollout of Astra to selected organizations on September 3, 2026.
Over the following days, the model is expected to become gradually available to users on the ChatGPT Plus, Pro, Business and Enterprise plans. Pro, Business and Enterprise users are also expected to receive access to the more powerful GPT-6 Astra Pro version.
The model is also expected to be available to developers through the OpenAI API and the Microsoft Azure and AWS Bedrock cloud services.
The rollout is gradual, so two users on the same plan may not receive access to Astra at the same time.
Will free ChatGPT users get GPT-6 Astra?
At the launch, OpenAI announced availability for Plus, Pro, Business and Enterprise. The company has not yet announced general access to GPT-6 Astra for the free plan.
That does not mean Astra or one of its derivative versions will never reach free users. At launch, however, OpenAI is focusing primarily on paid plans and professional products.
Is Astra simply a “smarter ChatGPT”?
Yes and no. For conventional questions, users may notice better reasoning, programming or handling of extensive context. However, the biggest change emerges only when Astra has access to tools.
The difference between older and newer generations of AI is increasingly unlike the difference between two chatbots. It is more like the difference between a system that explains a process and one that can carry out the entire process itself, check the result and then continue working.
That is precisely why OpenAI emphasizes the desktop application, ChatGPT Work, Codex and computer use when discussing Astra.
Is GPT-6 Astra a revolution?
After just one day of use and a set of benchmark results, it is too early to say that Astra is definitively changing the entire world of AI. Even exceptionally high test scores may not automatically translate into an equally dramatic difference for every everyday question.
Nevertheless, this represents an interesting shift in what OpenAI considers the core AI product. It is no longer merely about generating answers. The focus is increasingly moving toward the ability to use a computer independently, remember the course of complex work, create usable outputs and complete an entire task from initial instructions to final result.
In this respect, GPT-6 Astra may represent a more significant change than the number “6” in its name. If it works reliably outside carefully prepared demonstrations, ChatGPT will gradually evolve from a chatbot into a tool that can be entrusted with part of a real workload.
Sources
- OpenAI – GPT-6 Astra: A new generation of intelligence
https://openai.com/index/gpt-6-astra/ - OpenAI – Safety overview: GPT-6 Astra
https://openai.com/index/safety-overview-gpt-6-astra/ - OpenAI Developers – GPT-6 Astra Model
https://developers.openai.com/api/docs/models/gpt-6-astra - Reuters – OpenAI launches new Astra model amid growing scrutiny over agents' safety
https://www.reuters.com/legal/litigation/openai-launches-new-astra-model-amid-growing-scrutiny-over-agents-safety-2026-09-03/
Robert
I’m interested in technology and history, especially true crime stories. For three years I ran a fact-based portal about modern history, and for a year I co-built a blogging platform where I published dozens of analytical articles. I founded offpitch so that quality content wouldn’t be hidden behind a paywall.
