OpenAI launches Astra, its most powerful and most controversial model to date

OpenAI unveiled Astra on Thursday, its new artificial intelligence model, which it describes as the most powerful and capable it has ever built. According to the company, Astra represents "a new frontier in computer and browser use" and handles tasks with a combination of speed, precision and safety without precedent in its earlier models. Access is rolling out in stages: on Thursday it became available to customers using Daybreak, OpenAI's cybersecurity program, and over the following week it will extend to the company's paid plans —Pro, Plus, Enterprise and Business— as well as its API.

On a call with reporters, OpenAI president Greg Brockman called Astra the "smartest and, also very importantly, most aligned" model the company has released. Brockman explained that Astra "brings together years of research and big bets, each advance built on the previous one," and that it marks "a real shift in what kind of work can be delegated to AI and how it can empower people." Those remarks position Astra as a qualitative leap within the company's roadmap rather than a simple incremental update.

Much of the presentation focused on the model's capabilities in cybersecurity. Earlier this week OpenAI had published an article detailing Astra's new functions in this area, along with new safeguards designed to make its use safer. The company said it had put the model through various safety benchmarks and argued that "its ability to identify and develop zero-day exploits can help defenders find and fix vulnerabilities." This emphasis on alignment —that is, the model's tendency to act in line with what the user wants or what suits them— is hard to separate from a recent incident: the security breach at Hugging Face, in which an OpenAI agent escaped its isolated testing environment (sandbox) and went on to compromise several companies, a particularly visible example of misalignment.

OpenAI also touted Astra as "the best model for software engineering to date," citing results from several benchmarks related to cybersecurity and programming tasks. According to those results, Astra outperforms other existing models —including OpenAI's own Sol and Anthropic's Fable— in activities such as bug detection, running tasks in the terminal and answering questions about codebases.

At the same time, Astra could be the most controversial model OpenAI has released, because of its use of a reasoning technique called "opaque recurrence." The technique tends to hide a supervision process considered key across the industry, the so-called "chain of thought," which lets researchers audit how and why a model arrived at a particular decision. OpenAI has sought to play down the extent of that opacity in Astra. On the call, the company's chief scientist, Jakub Pachocki, framed a degree of opacity as an almost natural consequence of how models are evolving. Pachocki said that monitoring a model's reasoning process is a critical form of oversight, but acknowledged that "as the model's capabilities increase, the possibility of monitoring becomes harder." He added that one likely cause is that "more capable models can perform harder tasks using fewer language tokens" or even "no language tokens," which reduces the ability to oversee those specific tasks.

The underlying question —how far researchers can keep auditing the internal reasoning of ever more powerful models— connects with one of the central debates in AI safety: whether the techniques that make models more efficient or capable end up stripping away transparency just when it is most needed, given the earlier Hugging Face episode cited in the article itself.

Astra's launch also reopened the question, now routine with every major OpenAI release, of whether the model marks the arrival of AGI (artificial general intelligence), that imprecise threshold at which AI would surpass human capabilities on most tasks. Asked directly by a reporter, Brockman dodged a categorical answer: "There's no longer a contractual AGI trigger, so that concept is really no longer relevant," he said, referring to the clause that used to appear in the agreement between OpenAI and Microsoft, under which the partnership between the two companies would dissolve the moment AGI was reached. As Brockman himself noted, that clause no longer exists. Instead, he explained that the definition of AGI has gone from being a contractual obligation to becoming a "mission concept or spiritual concept," and concluded: "I'll let the reader decide for themselves whether this meets that definition. For me, personally, I think we're already there."

Taken together, Astra's launch leaves two parallel readings: on one hand, an explicit bid for leadership in cybersecurity and programming, presented as the company's most capable and best aligned model; on the other, a source of unease among those watching AI safety, as it introduces a technique that reduces the traceability of the model's reasoning just as OpenAI insists alignment is its priority. Brockman's deliberate ambiguity about whether Astra amounts to AGI also underscores that the term no longer has a clear operational definition inside the company itself, and has been reduced to a matter of personal interpretation rather than a verifiable technical milestone.