Greg Brockman has a habit of making bold claims about the state of artificial intelligence, and he made another one on the call announcing OpenAI’s newest model. Asked whether the company’s release of GPT-6 Astra represented a step toward artificial general intelligence, the OpenAI president said it no longer seemed unreasonable to describe the present as the AGI era. The remark, made on a conference call with reporters, was the most direct statement yet from a senior OpenAI executive that the company believes it has crossed a line it spent years describing as distant.
The model behind the claim is GPT-6 Astra, released Wednesday, which OpenAI says is its most intelligent and most alignment-focused system to date. The company describes improvements across the areas it has been emphasizing for the past year: operating computers, writing code, working through cybersecurity problems and conducting research. The release came almost exactly a year after GPT-5, and the compressed timeline is itself a statement about how OpenAI intends to compete: not with one flagship every few years, but with a steady drumbeat of releases that keep its name attached to the frontier.
The rollout is deliberately staged. Astra is available first to customers of Daybreak, OpenAI’s cybersecurity program, and will spread over the following week to the company’s Pro, Plus, Enterprise and Business plans and to the application-programming interface. Free users will have to wait. The staging reflects both technical caution and commercial strategy: the most demanding customers get the model first, and the broader market gets it once the early deployments have absorbed the initial problems.
The most consequential words in the announcement came from the company’s chief scientist, Jakub Pachocki. He acknowledged that Astra relies on what he described as opaque looped reasoning, a mode in which the model’s internal reasoning is not fully visible to the people trying to audit it. The admission is unusual for a company that has positioned safety as central to its mission: the more powerful the model, the harder it is for outsiders to verify that its reasoning is sound, and Astra’s design makes that verification problem worse rather than better.
The tension between capability and transparency runs through the entire release. OpenAI says Astra’s gains in computer use and cybersecurity come from letting the model reason for longer and with less external supervision, and those are precisely the settings that make audit difficult. Critics have argued that a model whose reasoning cannot be inspected is a model whose errors cannot be understood, and the company’s own chief scientist appears to accept the trade, betting that the capability gains outweigh the visibility costs.
The same evening, OpenAI announced a program designed to answer the safety questions its models raise. The company said it would commit $1 billion to Daybreak, an initiative to provide free AI security tools to critical infrastructure, including hospitals and power grids. The program positions OpenAI as a defender of the systems that artificial intelligence might someday threaten, and it gives the company a concrete answer to the question of what it is doing about the risks its own technology creates.
The $1 billion commitment is also a competitive move. OpenAI’s rivals have spent heavily on safety research and on partnerships with critical-infrastructure operators, and the Daybreak program puts OpenAI at the center of a conversation about AI security that regulators are increasingly driving. By giving tools away to hospitals and utilities, the company builds relationships with institutions that governments care about, and it does so at a scale its competitors will have to match.
Microsoft’s role in the rollout shows how tightly the two companies are bound. Azure, Microsoft’s cloud, began making Astra available to its enterprise customers within days of the announcement, extending the model’s reach into the corporate market that OpenAI reaches only through its own channels. The arrangement benefits both companies: Microsoft sells cloud capacity on the strength of OpenAI’s models, and OpenAI gains distribution it could not build alone.
The release lands at a moment when the industry is asking hard questions about the pace OpenAI has set. The company has released flagship models faster than any of its rivals, and each release has been accompanied by claims of new capabilities and new alignment techniques. The skeptics’ case is that speed and safety pull in opposite directions, and that a company releasing models annually will eventually release one before it is ready. OpenAI’s answer is that the alternative, waiting longer between releases, would cede the frontier to competitors with fewer scruples.
For now, the market has responded to Astra with the enthusiasm that has greeted every recent OpenAI release: enterprise customers are queuing for access, and the company’s partners are integrating the model into their products. Whether Astra justifies the description of the AGI era will be settled by use, not by announcements. The model will be tested in the places where OpenAI says it improved, in computer operation, in code, in security work and in research, and the results will determine whether Mr. Brockman’s remark was a prediction or a boast.
