OpenAI has officially flipped the switch on GPT-6 Astra, making it available to every Pro, Enterprise, and Business Premium subscriber directly within ChatGPT Work and Codex, with API access also live as of today. The announcement, which follows a phased rollout beginning September 3rd, means that only Plus and standard Business users will need to wait a few more days before they can access the new model.
The first piece of advice from OpenAI engineer Victor Nunez regarding this massive release? Delete your prompts. As Astra begins its rollout, he is urging developers to clean up their AGENTS.md files and Skills directories, and to rethink how they approach reasoning levels. AGENTS.md is essentially an employee handbook for AI, outlining who you are, what you should do, and what you shouldn't do. Skills, on the other hand, are instruction manuals teaching the model how to perform tasks, detailing specific steps for specific situations.
For years, developers have dealt with model limitations using a simple, blunt strategy: constantly adding more rules. If the model doesn't understand, add more explanation. If it makes mistakes, add more restrictions. If it still errs, add another example. It's like leaving sticky notes for a forgetful new employee, eventually covering the entire desk with reminders. With Astra, however, this mountain of sticky notes has become a liability. According to OpenAI's model guidance, Astra is far more sensitive to instructions in Skills and AGENTS.md files. Previously ignored rules are now executed one by one. A vague instruction will cause it to stop and ask for clarification, while two conflicting rules will leave it completely confused.
For example, a rule like "all plans require approval" was previously glossed over by older models. Astra, however, will literally stop and ask, "Who should approve this plan?" and then simply wait. Therefore, the guidance strongly recommends that developers audit every single Skill file the model can access. Furthermore, Astra is more inquisitive than its predecessors. It might stop mid-task to ask, "There are two ways to approach this; which do you prefer?" It will run tests before writing code and even demand unit tests for changing a button's color. All of this requires adjustment via prompts, and the direction is mostly about pulling back. The guidance even includes a ready-made prompt to prevent Astra from using clichés like "delve" or "it's worth noting," as well as banning the "not about X, but about Y" rhetorical reversal.
It's a fitting irony: the same company that created models prone to these phrases is now teaching you how to avoid them. It's an honest case of "I dug the hole, let me show you how to fill it." You might wonder how this affects you. The answer is significant. Astra's logic has changed, meaning your interaction style must change too. Previously, a poor response from ChatGPT would trigger a typical reaction to add more constraints. That’s additive thinking: the dumber the model, the more rules you write. Astra is telling you to stop writing so much. It can infer intent from context, ask questions when instructions are ambiguous, and remember overarching goals across multi-step tasks. Your patchwork prompts are not just redundant; they can lead it astray.
The recurring theme in OpenAI's official migration guide is "audit." Audit every instruction you give it, and delete anything outdated, redundant, or conflicting. The stronger the model, the fewer rules you should write. This principle applies to both developers and regular users alike. So what exactly has Astra gained? Early access developers have already provided answers. Developer Matt Shumer used Astra to build a Manhattan in Unreal Engine, street by street, over a week. He developed a method he calls the "manager loop": one Astra acts as a manager, breaking the task into checklists and phases, while another acts as an executor. The manager assigns one piece of work at a time, never revealing the bigger picture.
On the execution side, up to 96 sub-agents worked simultaneously, pushing the project forward like an assembly line. Shumer discovered a fascinating detail in wording: instructing the model to make each phase "excellent" kept it moving forward, but changing the word to "perfect" made it dive into minute details and get stuck. A single word distinction determined whether the project advanced or stalled. Developer Anshu exclaimed that Astra is a "super AGI machine god" for 3D gaming. Astra built a stunning open game world in just 45 minutes. Immunologist Derya Unutmaz simply entered "make a 5-minute educational video about T cells." Astra wrote the narration, used Remotion for animation, and Imagegen for graphics, and even suggested HeyGen for voiceover, producing the video in one shot. After 35 years of researching T cells, Unutmaz admitted he couldn't have explained it better himself and now plans to create a full immunology video course for his website.
Tom Krcha gave Astra an old blueprint of a steam locomotive. Minutes later, Blender produced 3,295 editable objects, each part individually customizable. This tweet has already received over 750,000 views. He then tried a more obscure model, the Commodore Vanderbilt train, which required manual adjustments to the front curvature, but the starting point was already impressive. OpenAI's own metrics are equally compelling. On OSWorld 2.0, which tests a model's ability to operate a computer desktop like a human, Astra achieved 72.6%, up from GPT-5.6 Sol's 65.7%. Task completion time dropped from 75 minutes to 40 minutes, meaning greater accuracy at nearly double the speed.
Of course, Astra doesn't dominate across the board. On Artificial Analysis' independent composite intelligence index, Astra scored 61.2 points, while Anthropic's Claude Fable 5.1, released this week, scored 65.7. The rivalry remains undecided. The past few years of AI development were about addition: compensating for a lack of intelligence with layers of rules. Now that those shortcomings are addressed, the rules themselves have become a source of errors. OpenAI's model guidance quietly redirects AI engineering: instead of teaching it every step, it's about dismantling the old, obstructive rules and giving it space. The relationship between humans and models has shifted. It's no longer about teaching it. It's about getting out of its way.
Comments