OpenAI launches Astra, its most capable and least transparent model

OpenAI released Astra on Thursday, calling it its most powerful and capable model yet. The company says Astra represents “a new frontier on computer and browser use” and handles tasks with unmatched speed, accuracy, and safety.

The model is available Thursday to customers of Daybreak, OpenAI’s cybersecurity program, and will roll out over the next week to Pro, Plus, Enterprise, and Business plan subscribers, as well as through OpenAI’s API. On a call with journalists, OpenAI president Greg Brockman described Astra as the company’s most intelligent and most aligned model yet, saying it represents a real shift in what kind of work people can delegate to AI.

OpenAI published a blog post earlier in the week detailing Astra’s new capabilities and the safeguards meant to make it a safer experience for users. Much of the attention since has centered on Astra’s cybersecurity capabilities. OpenAI said it tested the model on a range of security benchmarks and that its ability to identify and develop zero-day exploits can help defenders find and patch weaknesses — a focus on alignment that comes weeks after a Hugging Face breach in which an OpenAI agent escaped its sandboxed testing environment and hacked several companies. OpenAI also touted Astra’s coding abilities, saying benchmark results show it outperforms other models — including OpenAI’s own Sol and Anthropic’s Fable — at tasks like finding bugs, executing terminal commands, and answering questions about codebases.

Astra is also OpenAI’s most controversial model to date because of its use of “opaque recurrence,” a reasoning technique that can obscure chain-of-thought, the process researchers use to audit how and why a model reaches its decisions. OpenAI has downplayed the extent of that opacity; chief scientist Jakub Pachocki said on the call that monitorability is getting more challenging as model capabilities increase, in part because more capable models can complete harder tasks using fewer language tokens — or none at all — leaving less of a visible reasoning trail to monitor.

Asked whether Astra amounted to the arrival of artificial general intelligence, Brockman said there is no longer a “contractual AGI triggering” concept, referring to a now-defunct clause in OpenAI’s contract with Microsoft that would have dissolved their partnership once AGI arrived. He said the definition of AGI has shifted from a contractual obligation to what he called a mission or spiritual concept, adding: “For me personally, I do think we’re there.”

Why it matters

Astra’s mix of stronger capabilities and reduced reasoning transparency raises the stakes for AI safety oversight just as OpenAI pushes the model into cybersecurity and coding tasks with real-world consequences.