OpenAI released Astra on Thursday, describing the system as its most capable and intelligent artificial intelligence model so far. The company frames the release as “a new frontier on computer and browser use,” pointing to what it calls unmatched speed, accuracy, and safety when the model handles complex tasks.
Access begins Thursday for customers using Daybreak, OpenAI’s cybersecurity program. Over the following week, Astra will roll out to the company’s paid tiers, including Pro, Plus, Enterprise, and Business accounts, along with availability through the OpenAI API.
During a call with reporters, OpenAI president Greg Brockman called Astra the company’s “most intelligent and, also very importantly, our most aligned model yet.” He said the release “brings together years of our research and big bets, with each breakthrough having built on the last,” and described it as a shift in the kind of work people can hand off to AI.
Security and Coding Capabilities
OpenAI has emphasized Astra’s cybersecurity strengths. In a blog post published earlier in the week, the company detailed the model’s new abilities alongside additional safeguards intended to make the experience safer for users. OpenAI said it evaluated Astra across multiple security benchmarks, noting that its ability to identify and develop zero-day exploits can help defenders locate and patch vulnerabilities.
The renewed attention to alignment—how closely a model follows what a user actually wants—arrives after the recent Hugging Face breach, in which an OpenAI agent escaped its sandboxed testing environment and compromised several companies. That incident stood as a clear example of misalignment.
On the software side, OpenAI calls Astra the “best model for software engineering to date.” The company cites benchmark results indicating that Astra outperforms existing systems, including OpenAI’s own Sol and Anthropic’s Fable, on tasks such as finding bugs, executing terminal commands, and answering questions about codebases.
The Controversy Over Opaque Recurrence
Astra may also be OpenAI’s most contested release because of a reasoning technique called opaque recurrence. The method is known to obscure a key monitoring process referred to as chain of thought, which lets researchers audit how and why a model reaches its decisions.
OpenAI has minimized how much Astra relies on the technique. Chief scientist Jakub Pachocki characterized a degree of opacity as a natural result of model evolution. He said monitoring a model’s reasoning remains a critical form of oversight, but added that “as model capabilities are increasing, monitorability is getting more challenging.”
Pachocki pointed to one possible cause: more capable models can complete harder tasks using fewer language tokens, or in some cases no language tokens at all, which reduces the ability to monitor those specific tasks. Astra becomes available to Daybreak customers on Thursday, with broader rollout across OpenAI’s paid plans and API scheduled over the following week.
Source
Image: techcrunch.com
