OpenAI released GPT-6 Astra on Sept. 3, introducing its most capable model first to a limited group of trusted organizations after the system became the company’s first to reach its “Critical” cybersecurity threshold. A broader rollout to paid ChatGPT plans, the OpenAI API, Microsoft Azure and Amazon Web Services is planned over the coming days.
The launch matters because Astra is designed to act inside software, browsers and technical systems, not merely answer questions. That turns a model upgrade into an operating decision for businesses: The same capabilities that could accelerate software development, research and complex knowledge work also require tighter controls over permissions, monitoring and who can use the system.
From assistant to execution layer
OpenAI describes Astra as a step forward in computer use, browsing, software engineering, science and professional work. The company says the model can complete longer, multistep workflows while using fewer tokens than its previous flagship system on some evaluations.
For enterprise buyers, the practical shift is from generating recommendations to carrying out work. A model that can navigate applications, manipulate files and coordinate tasks may reduce handoffs across workflows, but it also expands the consequences of a mistaken instruction or excessive access. Deployment therefore depends as much on identity, authorization and auditability as on benchmark performance.
OpenAI set standard API pricing at $10 per million input tokens and $50 per million output tokens. Eligible API customers can use zero-data-retention controls, and enterprise administrators must actively enable Astra because access is off by default at launch.
Cyber capability becomes a deployment constraint
The most consequential part of the release is Astra’s cybersecurity capability. In its system card, OpenAI says the model can, with the right tools and access, find previously unknown flaws and develop new exploitation methods against well-protected systems without continuous human direction.
OpenAI is initially limiting advanced cyber workflows and says Astra will refuse some requests to produce proof-of-concept exploits. The company has added stronger monitoring, internal isolation and automated checks intended to stop unauthorized activity. Those safeguards can interrupt legitimate work, a trade-off OpenAI acknowledges.
The safety documentation also identifies a less obvious limitation: Astra’s written reasoning was harder to monitor than GPT-5.6 Sol’s in tests that asked the model to evade oversight. OpenAI says the model still struggled to conceal reasoning required for complex tasks, but called the decline in monitorability a research priority.
Governance has to move upstream
TechCrunch and Axios independently reported the launch and its unusually prominent safety questions. OpenAI President Greg Brockman said he believes the model may represent artificial general intelligence, but that is a company executive’s interpretation, not an independently established technical fact.
Executives evaluating Astra should treat the model as privileged infrastructure. Early deployments need narrow scopes, verified data boundaries, action logs, escalation paths and a clear human owner for consequential decisions. Procurement teams will also need to distinguish between access to Astra and permission for Astra to act across production systems.
The competitive significance is immediate even before broad availability. OpenAI is asking businesses to judge a frontier model on both the work it can execute and the controls surrounding that execution. Astra’s rollout makes those two questions inseparable.