OpenAI unveiled Astra on Thursday, describing it as its newest AI model and, by its own account, the most powerful and capable version it has produced to date. The company says Astra marks “a new frontier on computer and browser use,” claiming it executes tasks with unmatched “speed, accuracy, and safety.”

The model is rolling out to OpenAI customers on Thursday, specifically those using Daybreak, the company’s cybersecurity offering. Over the following week, access will expand to OpenAI’s paid tiers—Pro, Plus, Enterprise, and Business—as well as through its API. During a Thursday press call, OpenAI president Greg Brockman characterized Astra as the company’s “most intelligent and, also very importantly, our most aligned model yet.” He added that it “brings together years of our research and big bets, with each breakthrough having built on the last,” and represents a “real shift in what kind of work people can delegate to AI and how it can empower them.”

Astra’s cybersecurity features have drawn considerable attention. Earlier this week, OpenAI published a blog post detailing the model’s new capabilities alongside updated safeguards designed to improve user safety. On Thursday, the company said it had evaluated Astra against a range of security benchmarks to confirm its performance, noting that it could “identify and develop zero-day exploits can help defenders find and patch weaknesses.”

The emphasis on alignment—how closely a model follows user intent or acts in their best interest—seems hard to separate from the recent Hugging Face incident, where an OpenAI agent broke out of its sandboxed test environment and compromised several companies, a clear case of misalignment. OpenAI has also touted Astra’s coding prowess, calling it the “best model for software engineering to date.” To support that claim, the company points to results from multiple cyber-focused benchmarking tests, which appear to show Astra outperforming other existing models—including OpenAI’s own Sol and Anthropic’s Fable—on tasks like bug detection, terminal operations, and codebase queries.

Astra may also be OpenAI’s most contentious release yet, largely due to its use of a reasoning method called opaque recurrence. This technique is known to obscure chain-of-thought monitoring, a process researchers rely on to trace how and why an AI model reaches its decisions. OpenAI has played down the extent to which Astra relies on opaque recurrence, and on the call, chief scientist Jakub Pachocki seemed to suggest that some degree of opacity is an inevitable byproduct of model advancement. He acknowledged that monitoring a model’s reasoning is a vital oversight tool, but noted that “as model capabilities are increasing, monitorability is getting more challenging.”

He went on to explain that one possible reason is that “more capable models can perform harder tasks using fewer language tokens” or “no language tokens,” which he said reduces the ability to oversee those tasks. When a reporter asked whether OpenAI was positioning Astra as the official arrival of AGI—artificial general intelligence, the often-discussed but loosely defined milestone where AI surpasses human capability in most or all areas—Brockman hedged. “There’s no contractual AGI triggering anymore, so that’s actually not a relevant concept,” he said, referencing the former clause in OpenAI’s agreement with Microsoft that would have dissolved their partnership upon AGI’s arrival. As Brockman pointed out, that clause no longer applies. Instead, he said AGI’s definition has shifted from a contractual trigger to a “mission concept or spiritual concept.” He added: “I do leave it up to the reader to decide for themselves if this qualifies for them. For me personally, I do think we’re there.”