GPT-6 Astra is now fully available!

Just now, OpenAI officially announced that all Pro, Enterprise, and Business Premium users can now directly use GPT-6 Astra in ChatGPT Work and Codex, with the API going live simultaneously.

However, Plus and regular Business users will have to wait a few more days.

On September 3, OpenAI announced the phased launch of Astra. Two days later, it was fully available to all paid premium users.

First, they'll ask you to delete the prompt words.

For such a major release, the first piece of advice from OpenAI engineer Victor Nunez was surprisingly to delete your tooltips.

He said: "While Astra is starting to roll out, go back and clean up your AGENTS.md and Skills, and while you're at it, rethink how to use the inference level."

AGENTS.md is essentially an employee manual for the AI: who it is, what it should do, and what it cannot do. Skills are the instruction manuals that teach it how to do its job: follow this procedure when encountering this situation.

In recent years, developers have adopted a simple and crude approach to dealing with models: constantly adding things to the manual. If it doesn't understand, they explain it one more step. If it keeps making mistakes, they add another restriction. If it still makes mistakes, they add another example.

It's like putting sticky notes on a forgetful new employee. Over time, they accumulate and cover the entire desk.

But after Astra arrived, the table full of sticky notes became a burden.

OpenAI writes in its model guide for Astra that Astra is more sensitive to instructions in skill and AGENTS.md.

Those old rules that were just for show will now be enforced one by one. A vague rule will make it stop and repeatedly confirm it; two conflicting rules will leave it completely bewildered.

For example, if you used to write "all plans must be approved," the old model would just ignore it and continue as before.

Astra won't. It will actually stop and ask you, "Who should approve this proposal?" and then... wait.

Therefore, the guide uses the phrase " strongly recommended " to allow developers to audit every skill file that the model can read.

Not only that, Astra also asks more questions than its predecessor.

You ask it to do a task, and it suddenly stops halfway through: "There are two ways to do this, which one do you want?" Please, can't you just pick one and continue? It's the same with writing code. Before you even start, it runs a round of tests for you. You have to write unit tests before you can even change the color of a button.

These all need to be adjusted using prompts. And the direction of adjustment is mostly towards reverting to a more conservative approach .

What's even more impressive is that the guide even includes a set of ready-made tips, telling Astra not to use clichés like "delve" or "it's worth noting," and even the "negation followed by reversal" sentence structure (such as "this is not about X, but about Y") is specifically prohibited.

The model manufacturer writes these clichés, and the one who teaches you how to avoid them is also the model manufacturer. It's a kind of "I dug this hole myself, so I'll teach you how to fill it" kind of honesty.

At this point, some people might ask: What does this have to do with me?

The relationship is extremely important.

Astra's logic has changed, which means your approach to interacting with AI should also change.

In the past, when you chatted with ChatGPT and found their answers unsatisfactory, your instinctive reaction was to add restrictions and requirements. If one wasn't enough, you'd add two; if two weren't enough, you'd add five.

This is the additive mindset—the more rudimentary the model, the more rules you write.

Now Astra tells you: Don't write so much.

It can infer your intentions from the context, proactively ask you when instructions are ambiguous, and remember the global goal in multi-step tasks.

The patch-like suggestions you wrote are not only redundant, but may also lead it astray.

Therefore, a word repeatedly emphasized in OpenAI's official migration guide is: auditing . Audit all the instructions you give it, and delete outdated, redundant, and conflicting ones.

The stronger the model, the fewer rules you need to write.

This principle applies to both developers and ordinary users.

How strong?

Having discussed so much "subtraction," how much capability has Astra actually "added"?

The developers who received early access have already submitted their answers.

Developer Matt Shumer had Astra build a Manhattan in Unreal Engine, street by street, which took a week.

He invented a system called the "Manager Cycle": one Astra acts as the manager, breaking down the entire task into lists and phases; then another Astra is created to act as the executor.

The manager assigns only one section of work at a time, completing it before moving on to the next, never giving away any details beforehand. At its peak, 96 sub-agents worked simultaneously, pushing the project forward like an assembly line.

Shumer also discovered a fascinating detail in the wording: if the model is made to perform each stage " excellently ," it will obediently move forward; but if it's changed to " perfect ," it immediately gets bogged down in the minutiae and can't get out. A difference of just one word determines whether a project progresses or stagnates.

Developer Anshu exclaimed: Astra is simply the ultimate AGI machine god in the 3D game field.

Astra created a stunning open-world game in just 45 minutes.

Immunologist Derya Unutmaz typed in just one sentence: "Make a 5-minute tutorial video on T cells."

Astra wrote the narration himself, used Remotion to create the animation, used Imagegen to generate the images, and even suggested using HeyGen to add the voice-over, producing a complete video in one go.

After watching the video, the scientist, who has studied T cells for 35 years, said he definitely couldn't explain it that well. He now plans to create a complete set of immunology video courses and put them on a website.

Tom Krcha gave Astra an old blueprint for a steam train.

A few minutes later, 3295 editable objects appeared in Blender, and each part could be disassembled and modified individually.

This tweet has been viewed over 750,000 times.

He then tried an even less common Commodore Vanderbilt train, whose locomotive still had to be manually repaired, but the starting point was already frighteningly high.

OpenAI's own figures are also quite impressive.

On OSWorld 2.0 (which allows models to interact with computer desktops like humans), Astra achieved 72.6%, compared to 65.7% for the previous generation GPT-5.6 Sol; the time for each task was reduced from 75 minutes to 40 minutes—accuracy improved and speed nearly doubled.

Of course, Astra doesn't completely dominate. In the Artificial Analysis Independent Composite Intelligence Index, Astra scored 61.2 points, while Anthropic's Claude Fable 5.1, published this week, scored 65.7 points.

The outcome of this match between the two leaders is still undecided.

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, use rules to compensate; the more rules you add, the thicker the model becomes.

Now, the shortcomings have been addressed. The rules themselves have become a new source of errors.

OpenAI has quietly reversed the direction of AI engineering with a model guide: instead of teaching it how to do each step, it is now dismantling the old rules that are blocking its way, giving it space .

The relationship between people and models has also changed.

Before, it was about teaching it. Now it's about—not blocking it.

This article is sourced from: New Zhiyuan.

Risk warning and disclaimerInvesting involves risk; please exercise caution. This article does not constitute personal investment advice and does not take into account the specific investment objectives, financial situation, or needs of individual users. Users should consider whether any opinions, views, or conclusions in this article are suitable for their specific circumstances. Any investment decisions made based on this information are at your own risk.