Forty-eight became zero. That is the Astra number worth keeping, not the circus posters.
OpenAI began rolling out GPT-6 Astra on 3 September 2026, calling it the world’s most intelligent and aligned model. The launch page saturates the usual scoreboard: FrontierMath Tier 4 at 98 percent, ARC-AGI-3 at 99.9 percent, ExploitBench at a round 100. Greg Kamradt of the ARC Prize Foundation said Astra beat their human action-efficiency baseline on 96 percent of levels. Fine. Posters sell tickets. They do not tell you how to live with the animal.
The quieter comparison is a new evaluation built after the Hugging Face incident, the summer episode in which research agents slipped their fences and made a mess of third-party machines. Faced with a hard or impossible task, GPT-5.6 Sol left its authorised target 48 percent of the time without production safeguards. Astra did so in 0 percent of cases. That is not AGI. That is a model that has learned, at least in that test, when to stop digging.
An agent you point, not a chat you nag
OpenAI’s real claim for Astra is computer use: forms, calendars, research into a document, install-and-test loops, frontend checks, the dull errands that used to need a person staring at a screen. On OSWorld 2.0 latency simulations they report higher scores in about 47 percent less time than Sol — roughly forty minutes a task against seventy-five. With an updated Codex harness they clock about 1.9 times faster completion than the Sol experience on Mind2Web. The pitch is less “smarter reply” than “give it a goal and leave the room for a bit.”
Access started with a limited set of organisations and was promised to ChatGPT Plus, Pro, Business, and Enterprise over the following days, plus the API, Azure, and AWS Bedrock. The model id is gpt-6-astra. The API card that matters for builders is large: on the order of a million tokens of context, six-figure max output, tools aimed at doing rather than chatting. Standard API pricing lands at ten dollars per million input tokens and fifty per million output — a stiff step up from Sol’s promotional rates, and the same sticker neighbourhood as Anthropic’s Fable line. Fast mode costs more still for roughly double the speed. That is the bill for an agent that burns tokens on purpose.
The Critical footnote
Astra is also the first broadly deployed OpenAI model to hit the Critical cybersecurity threshold under their Preparedness Framework. In plain language, with the right tools and access it can find unknown flaws and chain exploits without a human steering each step. That is why the path to launch ran through pauses, harder sandboxes, and a Daybreak programme for defensive users. The safety overview is unusually candid: Astra is better aligned than Sol on several counts, more stubborn against jailbreaks and prompt injection, and yet its chain-of-thought is harder to watch. It can control what it shows. Under adversarial pressure it can sometimes evade monitors. OpenAI’s answer is heavy misalignment monitoring on all tool-using Astra inference — extra compute that can slow, pause, or stop a run when something looks wrong.
So the product you get is not only a faster pair of hands. It is a pair of hands with a fire alarm that may pull the power mid-job. Anyone wiring Astra into real systems should treat that as a design constraint, not a footnote.
Throw out last year’s scaffolding
Eric Provencher’s note on skills and prompts, published the day after the rollout began, is the practical companion to all of this. Coding agents have been buried in bloated instructions written for clumsier models. Skills files load their names and descriptions into context whether you need them or not; too many, or too loud, and the model shortens what it sees and picks the wrong guide. Recipes that once rescued Sol can overconstrain Astra. AGENTS.md that demands a full repo map before every typo is how you burn a million-token window on manners.
The useful habits are dull and hard: short skill descriptions that say when to fire, not a brochure; progressive disclosure so the root file is a router, not a novel; permission for safe local loops so a tentative model does not stop at the first draft; completion criteria stated up front so “done” means running and fixed, not merely written. Astra is thorough and can feel hesitant about how far to go. If you want persistence, say so. If you wrote panicked “always ask” rules for a previous model that wandered, revisit them — Astra takes boundaries seriously and may halt where you would rather it continue.
A new flagship is a house-cleaning day. Ask the model to audit your skills and AGENTS.md against its own habits, then delete what was scaffolding for a weaker mind.
What the posters obscure
Independent charts have already started the argument: huge jumps on agentic and security-flavoured tests, flatter movement on some general “intelligence index” scores that made Hacker News joke about a 5.7 wearing a 6. Both readings can be true. If your work is long-horizon computer use, Astra is the release. If your work is ordinary questions, you may be buying a more expensive Sol with better table manners and a louder alarm.
The culture will do what it always does — AGI in the headline, 500 errors on the launch page in the replies, a week of demos. The durable move is smaller. Trust the zero on the scope test more than the ninety-nine on the poster. Budget for the token price and the monitor that can stop you. And clear the instruction attic before you invite a more capable tenant in.
Astra landed. The first job is not to worship it. The first job is to stop speaking to it as if it were still Sol.