OpenAI has begun rolling out GPT-6 Astra, a model that company president and co-founder Greg Brockman says demonstrates artificial general intelligence has arrived. The release starts with business customers in the Daybreak programme, and a version with extra safety protections is planned for paid ChatGPT subscribers. Brockman’s statement is likely to fuel debate across the technology industry, in part because AGI remains one of the most contested terms in artificial intelligence research.
OpenAI defines AGI as systems that are generally smarter than humans. Brockman offered a more measured formulation when he explained the company’s position. “I do leave it up to the reader to decide for themselves if this qualifies for them,” he said. “I think we’re there.” That dual message, at once confident and open to interpretation, captures the unresolved tension surrounding OpenAI’s latest release.
What counts as AGI?
The term artificial general intelligence has never had a single universally accepted definition. Some researchers use it to describe a system that can perform any intellectual task that a human being can do. Others focus on the ability to transfer skills across very different domains without retraining. Still others argue that an AGI should be able to continuously learn from experience and operate autonomously in real-world settings. OpenAI’s working definition, based on outperforming humans across broad cognitive work, is stricter than some industry definitions but still leaves room for judgment calls.
Brockman’s claim is therefore not a purely technical statement. It is also a product and policy statement. By positioning Astra as the arrival point of AGI, OpenAI invites regulators, enterprise buyers and the wider public to treat the model as a significant threshold event. At the same time, by saying that readers can decide for themselves, Brockman acknowledges that no single benchmark can settle the question. The history of AI is filled with systems that seemed impressive in controlled tests but revealed limitations once deployed at scale.
Daybreak rollout and safety guardrails
GPT-6 Astra is being offered first to business users through OpenAI’s Daybreak programme. A separate version, designed for paid ChatGPT subscribers, will carry additional cybersecurity guardrails. Both versions restrict access to the most advanced offensive cyber capabilities. The company has said that Astra reached what it calls its critical cybersecurity threshold, a classification that has profound implications for how the model can be deployed and who is allowed to use it.
According to OpenAI, a model reaches that threshold when it can find and build working zero-day exploits against hardened systems without any human help. Zero-day exploits are vulnerabilities that software makers do not yet know about and therefore have not patched. If a model can identify such weaknesses and create reliable attack code by itself, the security risk is easy to understand. This classification is the same kind of internal threshold that led OpenAI to slow the release of a model in August, and it explains why the introduction of Astra took longer than some inside the company had hoped. Sam Altman, OpenAI’s chief executive, acknowledged that the release took longer than the team originally wanted.
OpenAI says Astra has been trained to use computers the way people do. Instead of simply answering questions in a chat window, Astra is designed to operate applications, move through interfaces and complete multi-step tasks. The company points to financial modelling, tax preparation and video game development as areas where the model can operate with minimal human oversight. Those capabilities are a major step beyond typical chatbot interactions, but they also introduce new risks. A system that can navigate software on its own could, if misused, cause serious financial or operational damage. That is why the decision to block the most powerful cyber features matters.
A voluntary national security preview
The U.S. government looked at Astra before its release. Brockman said that reviewers did not come back with demands to change the model. He offered that as reassurance, but the statement is weaker than it may appear. The June executive order on advanced artificial intelligence lets developers give federal agencies access for up to 30 days before a broad release. The order expressly says that nothing in it authorises mandatory licensing, preclearance or permitting. Participation in the review process is voluntary.
The director of the National Security Agency decides which models count as covered frontier models through a classified benchmarking process. That process is not a public assessment of safety or alignment. It is an intelligence classification designed to identify systems that may pose national security risks. Because the system is voluntary, no approval was withheld from OpenAI, but not because regulators concluded that Astra was completely safe. No approval was on offer in the first place. The distinction between passing a review and not being required to pass one is central to understanding the current oversight landscape.
Europe’s post-market approach
Europe does not pre-clear AI models either. The European Commission gained new powers on 2 August to conduct evaluations of advanced models. Those powers can be triggered by insufficient documentation or by an alert from the Commission’s scientific panel. Once an evaluation begins, regulators can demand access through application programming interfaces or through source code. These are meaningful enforcement tools, but they are post-market powers rather than gates that operate before a model reaches users.
Separately, from 11 September, manufacturers of products with digital elements must report actively exploited vulnerabilities within 24 hours of becoming aware of them. That reporting requirement creates a stronger accountability mechanism after incidents occur, but it does not create a pre-clearance process for frontier AI models. A company can still ship a model and only later face questions from regulators. The speed of AI deployment, combined with the slower rhythm of regulatory machinery, makes it difficult for governments to keep pace.
Internal safety framework under review
OpenAI has been rewriting its internal safety framework since the Hugging Face breach, a security incident that highlighted how vulnerable the AI supply chain can be. The company’s own classification system remains the binding constraint on releases like Astra. In the United States, the review process is limited to agencies getting early access. In Europe, the Commission can act after the fact. That leaves much of the responsibility with OpenAI’s internal decisions.
The challenge is that internal thresholds are not always visible to the public. OpenAI says Astra reached the critical cybersecurity threshold, but it has not published the full reasoning behind that determination. Classified benchmarking and proprietary evaluations create a layer of opacity that makes independent verification difficult. This is particularly concerning when a company simultaneously claims that its model meets the bar for artificial general intelligence and that it can protect against the most severe cyber risks. Regulators may want to know exactly how those two conclusions were reached.
The arrival of GPT-6 Astra therefore marks a moment of both progress and uncertainty. The model’s ability to handle complex tasks with minimal human oversight could reshape professional work and accelerate software development. But the same capabilities create dangerous possibilities if they fall into the wrong hands or are tested beyond their known limits. OpenAI’s decision to restrict access to advanced cyber functions is prudent, but it is not a complete answer to concerns about private and public sector misuse.
As other organisations work toward similar agentic systems, the questions raised by Astra will only become more pressing. What does it mean for an AI to be smarter than a human in a way that matters? How much trust should be placed in internal safety classifications that are not subject to outside review? And if governments decline to create mandatory pre-clearance systems, what other mechanisms can create accountability? For now, OpenAI’s own classification remains the binding one, and the burden falls on the company, its business customers and society to ensure that a model capable of reaching critical cyber thresholds is used with care.