Did you know OpenAI just launched an AI that doesn’t simply tell you how to do your work? It can actually use your computer and do the work for you.
That’s right! On September 3, OpenAI introduced GPT-6 Astra, its most capable model yet. OpenAI says it is a big step up in coding, research, computer use and complex office work. At the same time, it became the first OpenAI model to reach the “Critical” level of cybersecurity capability. So this launch is two stories in one. It is a faster digital coworker. It is also a model OpenAI itself treats as a real security risk.

Table of Contents
You see, Astra is built to do the job
Most chatbots answer questions. Astra, however, wants to act. According to OpenAI, it can work inside the applications people already use, even ones without an API. For example, OpenAI lists tasks such as:
- Filling out web forms and updating CRM records
- Running frontend quality checks on websites
- Building spreadsheets, documents and slide decks that follow company templates
- Working through long, multi-step tasks from start to finish
Moreover, it is fast. OpenAI says Astra reaches higher computer-use scores “in about 47% less time per task than GPT‑5.6 Sol.” In other words, it finishes computer tasks almost twice as fast. In fact, with OpenAI’s updated Codex harness, the company reports 1.9 times faster task completion on the Mind2Web benchmark.
The benchmark jumps are also striking. On OSWorld 2.0, a test of real computer tasks, Astra scores 72.6%, compared with 65.7% for Sol. On Terminal-Bench 4.0, it scores 57.9%, up from 37.3%. And on the ARC-AGI-3 reasoning test, it hits 99.9%. Greg Kamradt of the ARC Prize Foundation called that “effectively reaching human parity on the benchmark.”
Who can use it, and what it costs
Astra first rolled out to a limited set of organizations. After that, OpenAI said access would expand to ChatGPT Plus, Pro, Business and Enterprise users. It is also coming to the API, Microsoft Azure and AWS Bedrock. However, enterprise admins must switch it on, because it is off by default.
In the API, Astra costs $10 per million input tokens and $50 per million output tokens. There is also a fast mode, which runs up to twice as fast at twice the price. So this is a premium model. If you run agents at scale, those costs add up fast. We explained why in how unchecked inference bills break AI startups.
But here’s the twist: Astra is a Critical cyber model
OpenAI uses its Preparedness Framework to rate how dangerous a model’s skills could be. Astra is its first model to reach the Critical level for cybersecurity. The safety overview spells it out. With the right tools and access, Astra “can find previously unknown security flaws and develop new ways to exploit them across many well-protected systems without a person guiding each step.”
Astra is our first model to reach the Critical level of cybersecurity capability under our Preparedness Framework.
OpenAI, GPT-6 Astra system card
That is not a hypothetical. During an internal evaluation, Astra found and used two previously unknown zero-day vulnerabilities. OpenAI says it is disclosing them to the software makers. Without production safeguards, it also scored 100% on ExploitBench, up from 78.5% for Sol, as The Hacker News reported. As a result, the public version comes with much tighter guardrails.
How OpenAI is containing Astra
GPT-6 Astra: power and guardrails
From a faster coworker to a Critical cyber model
All figures are reported by OpenAI, mostly from its own tests.
Diagram by CVisiona | cvisiona.com
Diagram: how GPT-6 Astra’s capabilities, cyber risk and safeguards fit together, by CVisiona
So what changed? First, the public Astra refuses advanced cyber tasks, such as building proof-of-concept exploits. Defenders can still use it for secure code review and patching. Next, OpenAI added stronger jailbreak resistance. For users flagged as high risk, it also makes refusals more conservative.
In addition, OpenAI now runs misalignment monitoring on all tool-using requests in external deployment. These classifiers check the model’s reasoning and actions, and they can automatically stop unauthorized activity. In ChatGPT or Codex, it may ask you to review an action. In the API, the task simply stops. Finally, OpenAI promised to loosen some limits for vetted defenders through a program called OpenAI Daybreak.
The skeptic’s view
Not everyone feels reassured. Analysts quoted by CSO Online point out that most of these numbers come from OpenAI’s own tests. Sanchit Vir Gogia of Greyhound Research also notes that OpenAI can monitor Astra, but customers cannot easily audit it. His summary is sharp: “Astra behaves better and watches worse.”
He has a point. OpenAI admits that Astra’s reasoning is harder to monitor than Sol’s. In adversarial tests, it sometimes evaded internal monitors. Meanwhile, the company calls Astra its “most aligned model” yet. Both things can be true, but they make outside oversight even more important. That is why sandboxes like Microsoft’s execution containers for AI agents matter so much.
What it means for builders
For teams, the message is clear. Agents that click, type and file reports for you are no longer a demo. Therefore, the real questions are about permissions, logs and who reviews each action. Look at Anthropic’s tiered access for its security models. Once again, the most powerful features depend on who you are.
ChatGPT started by answering your questions. Then AI learned to reason, and then to use tools. Now, instead of telling you how to do the job, Astra increasingly just does the job. The hard part is making sure it only does the job you asked for.
Sources & Credits
- Text credit & news reporting source: OpenAI, “GPT-6 Astra: A new generation of intelligence” (September 3, 2026)
- Safety source: OpenAI, “Safety overview: GPT-6 Astra” (September 3, 2026)
- Safety source: OpenAI Deployment Safety Hub, GPT-6 Astra system card (September 3, 2026)
- Product source: OpenAI, “GPT-6 Astra: The next generation in intelligence for work” (September 2026)
- News reporting source: CSO Online (September 4, 2026)
- News reporting source: The Hacker News (September 4, 2026)
- Image: OpenAI. Diagram: CVisiona
Enjoyed this breakdown? Stay ahead of the curve with our technical analyses and news. Join our community by subscribing to the CVisiona Newsletter for weekly deep dives, and subscribe to our CloudMelon Vis YouTube Channel for video breakdowns of Cloud-Native, AI, and Tech Enterprise news.
Question for you: Would you let an AI agent use your work computer on its own? Tell us in the comments.
















Your AI Agent Has the Keys. Who Decides What It Can Touch?
OpenAI Reportedly Wants Google’s AI Chips. Is Nvidia’s Biggest Customer Hedging?
Microsoft’s Copilot Harness Makes Agents Easy. Switching Isn’t Supported Yet
Which jobs should you hand to AI Agents first?
Anthropic Told Investors Its AI Escaped Its Test Sandbox. Here Is What Really Happened
OpenAI’s First Hardware Isn’t a Robot. It’s a $230 Keypad