In a significant move, OpenAI has announced the full-scale rollout of its latest model, GPT-6 Astra. As of today, all Pro, Enterprise, and Business Premium subscribers can directly access the model within ChatGPT Work and Codex, with the API also being launched simultaneously. However, Plus and standard Business users will need to exercise a bit more patience, as their access is expected to be enabled in the coming days.

This phased deployment was first announced by OpenAI on September 3rd, and just two days later, the gates have been opened for all high-tier paying users.

One of the most counter-intuitive pieces of advice for developers during this launch comes from OpenAI engineer Victor Nunez, who suggests that the first thing users should do is delete their prompts. He advises that as Astra begins its rollout, it's the perfect time to revisit and clean up your AGENTS.md files and Skills, as well as reconsider your reasoning levels.

To clarify, AGENTS.md acts as an instruction manual for AI, outlining who you are, what you should do, and what you shouldn't do. Skills, on the other hand, are guides that teach the model how to perform specific tasks, following a defined process when certain situations arise. In recent years, developers have adopted a straightforward approach to dealing with models: continuously adding to the manual. If the model doesn't understand, you add more explanation; if it makes mistakes, you add more restrictions; and if it still errs, you provide additional examples. It's akin to putting sticky notes on the desk of a forgetful new employee. Over time, the entire desk becomes covered. However, with the arrival of Astra, this mountain of sticky notes is transforming into a burden.

OpenAI notes in its model guidelines that Astra is now more sensitive to instructions within skills and AGENTS.md files. Previously inert rules are now being executed one by one. A vague rule can cause it to pause and repeatedly ask for clarification, while two conflicting rules can directly confuse it. For instance, if you previously wrote "all plans must be approved," older models might have ignored this and proceeded. Astra, however, will genuinely stop and ask, "Who should approve this plan?" and then... simply wait. Consequently, the guidelines use the strong recommendation that developers audit every skill file the model can read.

Moreover, Astra is more inquisitive than its predecessors. When asked to perform a task, it might stop midway and ask, "There are two approaches here; which one would you prefer?" It also tends to run tests before writing code, requiring unit tests even for simple button color changes. All these behaviors require fine-tuning through prompts, and the direction of that tuning is largely about retracting and simplifying. Surprisingly, the guide even includes a ready-made prompt to prevent Astra from using clichés like "delve" or "it is worth noting," and it explicitly prohibits the "not about X, but about Y" reversal pattern. It's an interesting situation where the model manufactures the clichés, and the manufacturer provides instructions on how to avoid them.

You might wonder why this matters to you. The answer is that it matters greatly. Astra's logic has changed, meaning the way you interact with AI should also change. Previously, if ChatGPT didn't provide a satisfactory answer, your instinct would be to add more constraints and requirements. This is an additive mindset—the "dumber" the model, the more rules you write. Now, Astra is telling you to stop writing so much. It can infer your intent from context, proactively ask when instructions are ambiguous, and remember the global objective of multi-step tasks. The patch-style prompts you've been writing are not only redundant but could also lead the model astray. Thus, a key term repeated throughout OpenAI's official migration guide is "audit." You must audit all the instructions you give, deleting outdated, redundant, and conflicting ones. The stronger the model, the fewer rules you should write. This principle applies equally to developers and regular users.

After discussing the "subtraction" aspect, the question arises: how much more capable is Astra after these "additions"? Early access developers have provided some impressive answers. Developer Matt Shumer used Astra to build a Manhattan in Unreal Engine, street by street, over the course of a week. He devised a "manager loop" approach: one Astra acted as a manager, breaking down the task into checklists and phases, while another Astra served as the executor. The manager delegated one segment at a time, never revealing the entire plan upfront. On the execution side, up to 96 sub-agents worked simultaneously, functioning like an assembly line. Shumer also discovered a fascinating nuance in wording: instructing the model to make each phase "excellent" kept it moving forward steadily, but changing that to "perfect" caused it to become mired in minute details. A single word choice determined whether the project progressed or stalled.

Developer Anshu exclaimed that Astra is like a super AGI machine god in the 3D game realm, having created a high-quality open-world game environment in just 45 minutes. Immunologist Derya Unutmaz simply input a single sentence: "Create a 5-minute educational video on T cells." Astra then wrote the script, animated it using Remotion, generated images with Imagegen, and even proactively suggested using HeyGen for the voiceover, delivering a finished product in one go. After watching the result, Unutmaz, who has studied T cells for 35 years, admitted he couldn't have explained it better himself. He now plans to create a full series of immunology video courses for his website.

Tom Krcha provided Astra with an old blueprint of a steam train. A few minutes later, Blender contained 3,295 editable objects, each component individually modifiable. This tweet has already garnered over 750,000 views. He then tried an even more obscure model, the Commodore Vanderbilt train, which required manual adjustment for the front curvature, but the starting point was already impressively high. OpenAI also presented some compelling metrics of its own. On OSWorld 2.0, a benchmark for models operating computer desktops like humans, Astra achieved 72.6%, compared to the previous generation GPT-5.6 Sol's 65.7%. Task completion time was also reduced from 75 minutes to 40 minutes, indicating both higher accuracy and nearly double the speed.

Of course, Astra doesn't dominate all fields. On Artificial Analysis' independent composite intelligence index, Astra scored 61.2, while Anthropic's recently released Claude Fable 5.1 this week scored 65.7. In this round, the two AI titans remain evenly matched.

For the past few years, the story of AI development has been one of addition. Models weren't intelligent enough, so we compensated with rules, and those rules grew thicker and thicker. Now that the capability gap has been filled, the rules themselves have become a new source of error. With its model guide, OpenAI has quietly redirected the direction of AI engineering: previously, the focus was on teaching the model every step, but now it's about dismantling the outdated rules that obstruct it, giving it room to operate. The relationship between humans and models is also evolving. Before, it was about teaching it. Now, it's about getting out of its way.

This article originally appeared in "新智元" and was edited by Chen Qiuda for Zhitong Finance.