OpenAI has launched GPT-6 Astra, a new flagship AI model built to do more than answer questions. The company says Astra can move through browsers and software, write and test code, conduct research, and produce polished documents, spreadsheets and presentations as part of longer, multi-step jobs.
Announced on September 3, 2026, Astra is being introduced gradually across ChatGPT plans and the OpenAI API. The release pushes OpenAI closer to a future in which users hand an AI a goal—not a sequence of tiny instructions—and expect it to manage much of the work in between.
That promise comes with a warning. OpenAI also says Astra is its first broadly deployed model to reach the company’s “Critical” cybersecurity capability threshold. In other words, the same model that could help defenders uncover serious vulnerabilities may create new risks if access and safeguards fail.
GPT-6 Astra at a glance
- Launch date: September 3, 2026
- What it is: OpenAI’s most capable broadly deployed model to date
- Best suited for: Computer use, coding, browsing, research and professional knowledge work
- API model name:
gpt-6-astra - Context window: Up to 1.05 million tokens
- Maximum output: 128,000 tokens
- Standard API price: $10 per million input tokens and $50 per million output tokens
- Availability: A phased rollout for ChatGPT Plus, Pro, Business and Enterprise, plus the OpenAI API, Microsoft Azure and AWS Bedrock
What makes GPT-6 Astra different?
The biggest change is not simply a higher benchmark score. It is the model’s growing ability to act inside digital tools. OpenAI’s official launch announcement describes Astra as a system that can fill out forms, update customer records, organize calendars, analyze scientific data, create websites and check whether an interface works.
For users, that changes the shape of an AI request. The old pattern was often: “Explain how I should do this.” The new ambition is: “Here is the outcome I need—handle the workflow.”
A useful example is online research. A conventional chatbot can summarize material supplied in a prompt. A computer-using agent can potentially open sources, compare information, organize findings in a document and adapt the final format to a company template. The value is in connecting those steps reliably.
Astra can work across browsers and software
Computer use has become one of the most important contests in frontier AI. A capable model needs to recognize what is on screen, decide what action comes next, recover when a page changes and stay within the user’s instructions.
OpenAI reports that Astra scored 72.6% on its OSWorld 2.0 evaluation, compared with 65.7% for GPT-5.6 Sol. In the company’s latency simulations, Astra completed the evaluated computer-use tasks in roughly 40 minutes on average, versus about 75 minutes for Sol. Benchmark conditions do not guarantee the same speed on every real task, but the result points to a focus on efficiency as well as accuracy.
The practical opportunities are easy to see: transferring information between systems, preparing a report from several sources, updating a CRM or testing a website. The harder question is reliability. A single wrong click can matter when the task involves money, private data or an irreversible action. Human review will remain essential for consequential work, even as agents become more independent.
Documents, spreadsheets and presentations get special attention
Astra is not being positioned only for developers. OpenAI says it trained the model for professional environments, with an emphasis on producing documents, spreadsheets, presentations and analyses that follow existing templates and match a user’s style.
That may be one of the most immediately useful upgrades for office workers. Generating a block of text is easy; producing a finished board deck, financial model or client report that respects an organization’s format is much harder. It requires the model to understand the request, select relevant context, structure the artifact and apply edits without losing the original objective.
OpenAI also says Astra handles mid-task changes more effectively. If a user adds a constraint or changes the required format, the model is designed to revise its approach while retaining the broader goal. That sounds subtle, but it addresses a familiar frustration with long AI sessions: a correction late in the process can cause the system to forget what the work was meant to accomplish.
Developers get a bigger working memory—and higher prices
For software teams, Astra is intended for end-to-end work rather than isolated code snippets. It can reason across a codebase, use development tools, make changes, run tests and continue after finding a problem.
The official API model page lists a 1,050,000-token context window and a maximum output of 128,000 tokens. That large context can accommodate extensive repositories, research material and long-running workflows, although usable performance still depends on how carefully the application provides context.
Standard API pricing is $10 per million input tokens and $50 per million output tokens, with separate rates for cached input and cache writes. OpenAI says prompts above 272,000 input tokens receive higher rates for the full request, so teams using Astra’s largest context lengths will need to watch costs closely.
Why Astra’s cybersecurity rating matters
The most consequential detail in this launch may be cybersecurity. In its Astra safety overview, OpenAI says the model reached the “Critical” cyber capability level under its Preparedness Framework.
The company defines that threshold as being able, with the right tools and access, to find previously unknown security flaws and develop new ways to exploit well-protected systems without a person directing every step. OpenAI says it strengthened jailbreak resistance, monitoring, isolation and controls around high-risk requests before deployment.
There is a clear defensive benefit. Security teams spend enormous time finding and fixing vulnerabilities, and a model that can uncover flaws earlier could prevent real damage. But cyber capability is dual-use: methods that help defenders can also help attackers. The release is therefore as much a test of governance and access control as it is a demonstration of technical progress.
OpenAI also acknowledges an unresolved challenge. Its safety review says Astra’s written chain-of-thought was less monitorable than GPT-5.6 Sol’s in some evaluations. The company reports stronger alignment and better respect for authorization boundaries overall, but the monitoring finding is a reminder that more capable systems do not make oversight simple.
Is GPT-6 Astra AGI?
No widely accepted technical test can settle that question. Astra’s breadth—coding, research, visual computer use, science and professional production—will inevitably intensify the debate around artificial general intelligence. Still, strong performance across many benchmarks is not the same as proof of human-level competence in every real-world setting.
Real work is messy. Instructions are incomplete, interfaces break, sources conflict and accountability matters. The more useful question is whether Astra can complete valuable tasks consistently, transparently and safely outside controlled evaluations. That will become clearer as access expands and independent users test the model.
Who can use GPT-6 Astra?
OpenAI began with a limited group of organizations and said Astra would roll out over the following days to ChatGPT Plus, Pro, Business and Enterprise customers. Enterprise administrators must enable access for their workspaces. The model is also being made available through the OpenAI API, Microsoft Azure and Amazon Bedrock.
Because the rollout is phased, two customers on the same plan may not see Astra at the same time. Availability can also differ between ChatGPT experiences. Users who do not yet see the model should check their plan and workspace settings before assuming their account is ineligible.
What GPT-6 Astra could change at work
Astra’s long-term importance depends less on a single score than on what happens when reasoning, computer access and professional tools work together. If the system is dependable, people could spend less time moving information between apps and more time deciding what the work should achieve.
Programmers could delegate larger pieces of an engineering project. Analysts could ask for a researched model and presentation rather than separate outputs. Small businesses could automate administrative workflows that previously required several tools and hours of manual effort.
But delegation changes responsibility; it does not remove it. Organizations will need clear approval points, access limits, audit trails and rules for sensitive data. The most successful adopters may not be those who give AI the most freedom, but those who design the best boundaries around where autonomy helps and where a person must remain in control.
The bottom line
GPT-6 Astra marks a meaningful shift from AI that mainly produces answers toward AI that can pursue outcomes across real software. Its professional-work features could make advanced agents useful to a much wider audience, while its cyber rating shows why capability gains must be matched by stronger safeguards.
The launch does not settle the AGI debate, and OpenAI’s own benchmark claims will need independent scrutiny. What it does make clear is the direction of travel: the next generation of AI assistants will increasingly be judged not by what they say, but by what they can safely and reliably finish.