OpenAI began rolling out GPT-6 Astra on Wednesday to a limited set of organizations, with ChatGPT Plus, Pro, Business and Enterprise users, the API and AWS to follow over the coming days. Companies enrolled in its Daybreak cybersecurity program were served first.
It is the same model the company classified at the Critical cybersecurity level of its Preparedness Framework two days earlier, the first of its models to reach that tier, and OpenAI says the rating is why the rollout is staged.
The cyber results are the reason for the rating. OpenAI reports a perfect score on ExploitBench, which measures turning known vulnerabilities into working exploits, against 78.5 percent for GPT-5.6 Sol. On SRE-Bench, which tests reverse engineering of software binaries without source code, it says Astra solved 88.0 percent of tasks on the first attempt and 99.2 percent within four, where Sol managed 55.9 and 68.7.
The general claims are large and self-reported. OpenAI calls Astra “the world’s most intelligent and aligned model,” and says it saturates FrontierMath Tier 4 at 98 percent and ARC-AGI-3 at 99.9 percent. It puts Astra at 57.7 percent on Terminal Bench 4.0, just above the 55.8 percent Anthropic reported for Claude Fable 5.1 three days ago on the same benchmark, each company scoring its own model. Configurations differ between runs, and features such as persistent memory change results.
What the model is built for is computer use. OpenAI says ChatGPT is close to twice as fast at operating a desktop with Astra, and that harness optimizations shipped alongside it speed up GPT-5.6 Sol by around 60 percent. In Codex, Astra can keep notes across context windows instead of compacting a session into a single summary, with earlier windows left searchable; the feature is experimental now and becomes the default in the coming weeks. A company demonstration had the model doing 3D modeling, assembling slideshows and ordering food while coding a game.
API pricing is $10 per million input tokens and $50 per million output, the same rate Anthropic charges for Fable 5.1. Astra usage counts against existing subscription allowances, with additional credits purchasable, and Pro, Business and Enterprise plans also get a GPT-6 Astra Pro tier. Enterprise administrators have to switch it on, since access is off by default.
The staging caused its own problem. Pro subscribers, who normally see new models first, watched enterprise customers get ahead of them, and Altman apologised on Thursday for a rollout he called messy. Codex engineering lead Thibault Sottiaux offered affected users one banked usage reset for each day they wait, which applies to the models they already have rather than to Astra.
The release lands less than two months after the GPT-5.6 Sol, Terra and Luna family, and after a summer in which OpenAI slowed frontier work following the Hugging Face breach. Astra was itself paused in August, though it was not one of the models involved.
Sources: 9to5Mac, Dataconomy
–
By the Control Plane Editorial Team