OpenAI releases GPT-6 Astra, available to Pro, Enterprise, and Business Premium users, with API now live.Article author and source: AI New Era
GPT-6 Astra is now fully open!
Just now, OpenAI officially announced that all Pro, Enterprise, and Business Premium users can now directly use GPT-6 Astra within ChatGPT Work and Codex, with the API now live.
However, Plus and regular Business users will need to wait a few more days.
On September 3, OpenAI announced the phased rollout of Astra. Today, two days later, access has been fully opened to paid premium users.
First, ask you to delete the prompt.
For such a major release, OpenAI engineer Victor Nunez’s first piece of advice to everyone was to delete your prompt.
He said: As Astra is launching, take a moment to clean up your AGENTS.md and Skills, and reconsider how to use the reasoning levels.
AGENTS.md is the employee handbook for AI: who you are, what you should do, and what you shouldn't do. Skills is the instruction manual for performing tasks: follow this process when encountering this situation.
Over the past few years, developers have dealt with models in a straightforward, brute-force way: constantly adding more to the manual. If it doesn’t understand, explain one more step. If it keeps making mistakes, add another restriction. If it still errs, provide one more example.
Like putting sticky notes on a new employee who always forgets things—over time, the entire desk becomes covered in them.
But after Astra arrived, all these sticky notes on the table became a burden.
OpenAI wrote in Astra's model guidelines: Astra is now more sensitive to instructions in skill and AGENTS.md.
The old rules that were merely for show are now being enforced point by point. A vague rule will cause it to pause and repeatedly verify; two conflicting rules will simply confuse it.
For example, previously you wrote, “All plans must be approved,” but the old model ignored it and continued anyway.
Astra won’t. It will genuinely stop and ask you, “Who should approve this plan?” and then… wait.
Therefore, the guide uses the phrase "strongly recommends" to ensure that every skill file readable by the developer audit model is reviewed.
不仅如此，Astra 还比上一代更爱提问。
You ask it to do a task, and halfway through it stops: “There are two approaches here— which one do you want?” Come on, can’t you just pick one and keep going? Even when writing code, it runs a full test suite before you’ve even started—changing a button color requires writing unit tests first.
All of this must be adjusted through prompts, and the direction of adjustment is mostly toward tightening.
Even more striking, the guide included a ready-made prompt instructing Astra to avoid clichés like “delve” and “it is worth noting,” and specifically banned the “negate then flip” structure (e.g., “This isn’t about X, it’s about Y”).
The model wrote these clichés, and the same company that made the model teaches you how to avoid them—there’s a kind of honesty in “I dug the hole myself, and now I’m teaching you how to fill it.”
At this point, someone might ask: What does this have to do with me?
It's too significant.
Astra's logic has changed, which means the way you interact with AI should change too.
Previously, when you chatted with ChatGPT and found its answers lacking, your instinct was to add restrictions and requirements—one wasn’t enough, so you added two; two weren’t enough, so you added five.
This is the additive approach—the dumber the model, the more rules you write.
Now Astra tells you: Don't write so much.
It can infer your intent from context, proactively ask for clarification when instructions are ambiguous, and remember the overarching goal across multi-step tasks.
The patchwork prompts you wrote are not only unnecessary but may also lead it astray.
Therefore, the official OpenAI migration guide repeatedly emphasizes one word: audit. Audit all the instructions you provide it, and remove all outdated, redundant, or conflicting ones.
The stronger the model, the fewer rules you should write.
This principle applies equally to developers and ordinary users.
How strong?
After all this talk about "subtraction," just how many capabilities has Astra actually added?
Developers who received early access have already submitted their responses.
Developer Matt Shumer built a Manhattan in Unreal Engine, street by street, over the course of a week.
He invented a gameplay system called the "Manager Cycle": one Astra acts as the manager, breaking down the entire task into checklists and phases; another Astra is assigned as the executor.
The manager assigns tasks one segment at a time—only releasing the next segment once the current one is completed, with no advance previews. On the execution side, at peak capacity, up to 96 sub-agents work simultaneously, pushing tasks forward like an assembly line.
Shumer also discovered a fascinating nuance in phrasing: when instructing the model to achieve each stage to "excellence," it readily moves forward; but as soon as the wording is changed to "perfection," it immediately gets stuck in the minutiae. A single word’s difference determines whether the project progresses or stalls.
Developer Anshu exclaimed: Astra is literally the supreme AGI machine god of the 3D gaming world.
Astra created a stunningly high-quality open game world in just 45 minutes.
Immunologist Derya Unutmaz entered just one sentence—“Create a 5-minute tutorial video on T cells.”
Astra wrote the narration herself, created animations with Remotion, generated images with Imagegen, and proactively suggested using HeyGen for voiceover—all in one go.
After watching, the scientist who has studied T cells for 35 years said he couldn’t have explained it any better. He now plans to create a complete series of immunology video courses to post on his website.
Tom Krcha gave Astra an old blueprint of a steam locomotive.
A few minutes later, Blender produced 3,295 editable objects, each of which can be individually detached and modified.
This tweet has received over 750,000 views.
He also tried a rarer Commodore Vanderbilt locomotive, whose curved front required hand-finishing, but even its starting point was astonishingly high.
The figures OpenAI has disclosed are also impressive.
On OSWorld 2.0 (which enables models to operate a computer desktop like a human), Astra achieved 72.6%, compared to 65.7% for the previous generation, GPT-5.6 Sol; the time per task was reduced from 75 minutes to 40 minutes—accuracy improved while speed nearly doubled.
Of course, Astra is not superior in every aspect. On Artificial Analysis's independent composite intelligence index, Astra scores 61.2, while Anthropic's newly released Claude Fable 5.1 scores 65.7 this week.
This round between the two rivals has yet to be decided.
The Era of Subtraction
Over the past few years, the story of AI development has been one of addition.
If the model isn't smart enough, compensate with rules—the more rules you add, the thicker they become.
Now, the weakness has been addressed. The rule itself has become a new source of errors.
OpenAI quietly shifted the direction of AI engineering with a model guide: instead of teaching it how to do each step, they are now systematically removing outdated constraints to give it space.
The relationship between people and models has also changed.
Before, it was about teaching it. Now, it’s about not getting in its way.