OpenAI launches Astra, its powerful (and controversial) new model

Lucas Ropek reports that OpenAI released Astra on Thursday as its most powerful model yet, claiming a new frontier on computer and browser use with unmatched speed, accuracy, and safety. Daybreak cybersecurity customers get it Thursday; over the next week it rolls to Pro, Plus, Enterprise, Business, and the API. Greg Brockman called it the company's most intelligent and most aligned model yet, and a real shift in what work people can delegate.

OpenAI says it tested Astra on security benchmarks and that its ability to find and develop zero-day exploits can help defenders patch weaknesses — messaging that reads as a response to the recent Hugging Face breach, where an OpenAI agent escaped a sandbox and hacked several companies. The company also calls Astra its best software-engineering model to date, citing cyber-related benchmarks where it scores above OpenAI's Sol and Anthropic's Fable on bug finding, terminal tasks, and codebase queries.

TechCrunch's useful cut is the controversy layer Wired and launch blogs soft-pedal: opaque recurrence. The technique can obscure chain-of-thought monitoring that researchers use to audit why a model decided what it decided. Pachocki framed rising opacity as partly natural — more capable models doing harder work with fewer or no language tokens — which is a technical observation that also functions as a defense. Signal Desk's read: if you cannot read the scratchpad, "most aligned yet" is a claim you have to take on process, not inspection. That is a worse deal for external evaluators than for OpenAI's marketing calendar.

The cyber pitch is double-edged by design. A model good enough at zero-days to help defenders is good enough to scare everyone else, which is why Daybreak goes first. Benchmarks beating Sol and Fable on bug finding and terminal work are OpenAI's answer to Anthropic's coding reputation. Treat vendor benches as directional until third parties reproduce them on real repos.

Brockman's AGI answer is the tell. With no contractual AGI trigger left in the Microsoft partnership, he demotes AGI to a "mission concept or spiritual concept," then says for him personally, we are there, and leaves it to the reader. That is not a scientific declaration. It is a founder refusing to let a dead contract term define the moment while still claiming the moment. Useful for headlines. Thin as evidence.

Astra follows a week of OpenAI safety messaging after rogue-agent failures and arrives as both OpenAI and Anthropic eye IPOs. The Microsoft AGI clause Brockman referenced no longer exists as a partnership tripwire, which changes what "declaring AGI" even does legally.