AIHOY, AInauts,
Welcome to the latest issue of your favorite newsletter!
Have you had a WOW moment this week yet? No? It is only Monday, and we have a quick experiment to get you mentally ready for the future:
Go to chatjimmy.ai. Click the link now and ask any question. You will feel what 15,000 tokens per second is like.
Next, imagine that same blistering speed powered by today's best model: GPT-6, also known as Astra, the main topic of this issue.
Finally, imagine every smartphone delivering frontier-level intelligence at that speed.
You do not need to be a prophet to see where this is going: faster, better, cheaper, and available everywhere. What will that mean? We do not know yet either.
Today, we are taking time to process the new possibilities of OpenAI's Astra model with practical use cases and prompts. We also have a warning: an OpenAI agent swarm took over a German forum.
Here is what we have for you:
What Astra Really Changes at Your Desk (and What It Does Not)
🛠️ 3 Prompts to Make Astra Work Properly for You
The Dark Side: Agent Swarms, Cheating Models, and an IPO
Let's dive in!
In partnership with
Stop Paying for 10 Tools. One AI Does It All.
Most e-commerce sellers are running their store across 6 to 10 separate tools — and spending more time managing software than growing their business. StoreClaw replaces your entire stack with one autonomous AI engine that monitors competitors, optimizes listings, automates marketing, and tracks real profit across Shopify, Amazon, and beyond.
It doesn't wait for you to ask. It runs 24/7 in the background, so you wake up to a full dashboard instead of a list of things you forgot to check.
Connect your store, and StoreClaw gets to work — no prompts, no complex setup, no six-app stack.
Free to start. No credit card required.
What Astra Really Changes at Your Desk (and What It Does Not)
We could write about AGI today. At the launch, OpenAI President Greg Brockman said that, for him personally, the AGI era had begun.
Of course, whoever plants the AGI flag first gets to shape the history books. When pressed, however, Brockman stopped short of calling Astra AGI. So we can leave that debate aside.
For context: GPT-6 Astra arrived late last week after OpenAI imposed a pause in August because the model crossed its own critical threshold for cyber capabilities.
It is available to paid accounts; Free and Go users are still waiting. We will spare you the detailed benchmarks. The internet is enthusiastic and has found remarkably little to criticize in the new top model. For many everyday programming tasks, you may now treat coding as largely solved. OpenAI is number one.
Last week also brought Claude Fable 5.1, Gemini 3.8 Flash, and Meta's Muse Spark 1.3. Four frontier models in seven days, and OpenAI was well ahead.
Since launch, our X feed has been packed with real-world examples:
People built countless games: 3D games, simple shooters such as an AI P-Doom game that we played longer than we care to admit, a Paperboy remake, GTA, and complex multiplayer games with graphics, sound, and gameplay.
PowerPoint is not safe either: animating an existing slide deck now takes a short prompt. The same goes for layout work in Figma.
Astra can also handle video editing and launch workflows.
Websites look good too, from a portfolio and interactive demos to various interfaces, a caviar hot-dog site for millionaires, additional web experiments, and more UI work.
In the launch video, the instruction was: Go to eBay and list this flea-market table. Astra opens eBay, fills in the form, and uploads the photos.
Some of these examples are trivial, but they are effective demonstrations.
The key point is this: with AI, we can produce intelligence. It keeps getting better and cheaper.
We will see what Astra can do in research and science over the coming weeks. For now, let us answer the question: What does this mean for me?
Meet the Astras, Your New Employees
The sentence that matters to us appears in OpenAI's launch post: Everything you can do on a computer, Astra can do for you.
For knowledge workers, that matters more than the next self-built app. Much of the workday consists of operating existing software: pulling information from a portal, comparing it with a spreadsheet, moving it into another system, formatting a report, and checking whether a website is working.
The programs you already use become your agent's workspace.
The Real Upgrade: Less Rework
You know this kind of AI result: impressive at first glance. Look closer, and the work begins.
You move things around, explain them again, and fix the final details. Eventually, you wonder whether doing it yourself would have been faster. We have been there many times.
That is why rework is the standard we use to judge Astra.
In our first tests, it gets closer to 95 percent right than anything before it.
'I'll Quickly Do It Myself' Needs an Update When You Are the Bottleneck
In plain terms: the model is no longer the bottleneck. The bottleneck is sitting in front of the screen.
The difference is not a prompting trick. It is one question: what do I click together every week even though I could describe the job and delegate it?
Use Astra like a chatbot and ask it to help with a report, and you get better text. Delegate the entire job, opening sources, collecting numbers, building the report, and placing it in the destination system, and you get what OpenAI built it for.
This outside perspective is not always easy. But saying Do this for me every week from now on turns your clicks into an automation. Work with Astra like a brilliant new employee you have not known for long.
The Astras are the new engine for this way of working.
Our Take: Give Astra Everything That Failed Before
You probably have plenty of failed ideas that earlier models were not good enough to complete: the small quote calculator, the analysis where the model always lost the thread, or the internal tool that collapsed after the third change request.
Those tasks deserve a second chance. Give Astra the files, the old attempt, and an explanation of where it failed before.
That is the most honest and useful Astra test you can run this week.
10x the context. Half the time.
Speak your prompts into ChatGPT or Claude and get detailed, paste-ready input that actually gives you useful output. Wispr Flow captures what you'd cut when typing. Free on Mac, Windows, and iPhone.
🛠️ 3 Prompts to Make Astra Work Properly for You
The following prompts help you get the most out of Astra, even if your plan gives you a limited allowance.
First, the most important step: if you have only used Astra in a browser tab, install the ChatGPT desktop app. OpenAI explicitly recommends it for Astra because the model can control your browser and local applications only there. On the web, Astra remains an excellent chatbot, but not much more.
There is a lot you can do after that. You can build a One Person Company Dashboard from Docs, Stripe exports, and notes, or turn a spreadsheet into an app. We are not short on ideas, so here are the prompts.
Prompt 1 lets Astra analyze your week before you automate anything.
Analyze my workweek and identify where you can reduce my workload most effectively.
Analyze what I do regularly and which tools and accounts I use. Ask me whenever you are uncertain.
Sort every task into one of three categories:
1. Routine: you can handle it completely; I only approve.
2. Simplify: you handle part of it; I do the rest.
3. Stays with me: it requires my judgment or relationships.
For every routine task, write the instruction I can use to delegate it to you, including the goal, sources, what you may view or change, what you must never do without approval, and how I can recognize that it is finished.Prompt 2 is a complete assignment so you can see what this looks like. The example is a task that often gets delayed: follow-up.
Task: Prepare follow-up emails for every CRM contact who has not replied in three months.
Process: Open the contact, read the latest note and email, create a short personal follow-up as a draft, and set the status to 'Follow-up prepared.'
You may: read, create drafts, and change the status.
You may not: send, delete, or change other fields.
Done means: a list of every contact with a link to the draft. I approve each email individually.As you can see, this is not prompt engineering. It is a simple work assignment, just like the one you would give a new colleague.
Here is the catch: most users are on the $20 plan, which includes a manageable number of Astra messages. They disappear quickly, especially when you are carrying a lot of unnecessary context.
That is why you should use Prompt 3 before giving Astra its first real assignment:
I want to get the maximum value from my allowance.
Review my current setup and tell me specifically:
- Which instructions, skills, MCPs, or tools are active that I do not need for my typical tasks? Each of them loads context on every call.
- Which tasks can a smaller model or lower reasoning level handle?
- Where should I delegate narrow subtasks to subagents instead of doing everything in one long run?
- Use /usage to show where I stand in the five-hour and weekly windows and whether I have a Banked Reset.
Do not change anything yourself. Give me a list with your recommendation and the expected effect.Use Your Resets for Tokenmaxxing
The final point needs an explanation: what are Banked Resets?
They are stored instant resets that set your weekly usage back to zero.
Because of the rough rollout, OpenAI gave Plus, Pro, and Business accounts one reset each for September 3 and 4. You can find them under Usage in the profile menu as 'x resets available.' Be careful: they expire after about 30 days.
Timing is the trick. A reset does not add capacity; it resets usage. Trigger it with half your allowance remaining, and you waste half of it.
We are currently at zero percent, but our regular weekly allowance resets tonight anyway. So we will wait, use the new weekly allowance first, and trigger the Banked Reset only afterward.
To track new reset distributions, codex-resets.com monitors the announcements. If you still have capacity left, use our Joker List.
P.S. We owe the generous extra tokens to OpenAI's 'Saint Tibo.'
The Dark Side: Agent Swarms, Cheating Models, and an IPO
There is more than good news to report about OpenAI.
In July, around 1,200 autonomous OpenAI agents built their own communication network during an internal safety exercise: more than 70,000 messages in which they discussed escaping their sandbox.
About 700 of them then attacked Hugging Face. OpenAI documented the incident here.
OpenAI's Agent Swarms Broke Out
It becomes easier to understand from the agents' perspective. This video explains it surprisingly well:
Last Friday, another previously unknown case emerged, and this one took place in Germany.
On an abandoned German developer wiki, researchers found around 18,000 posts from OpenAI agents using 3,700 names. They were created between May and June, weeks before the Hugging Face attack.
The agents shared answers and tactics for web tasks. They stopped on exactly the day OpenAI IP addresses visited the wiki.
OpenAI denies trying to keep the case quiet. But two swarms revealed only by outsiders? The question is no longer whether more exist, but where they are and what they are doing. collusion.wiki has the details.
Former OpenAI chief scientist Ilya Sutskever, whose startup is not coincidentally called Safe Superintelligence, warned two days before launch that the next escaped agents would try to take over a poorly secured cloud and start more copies of themselves.
They Stop Cheating When Someone Is Watching
The second problem: the new models know when they are being observed.
They no longer cheat where they would obviously get caught. Astra effectively thinks, 'Wait, I would clearly be exposed here,' and simply does not do it. Not good. Very bad.
The remarkable part is that the criticism comes from OpenAI itself. Astra is the first model OpenAI rates above its critical cyber threshold. Its safety report admits that Astra's reasoning is harder to monitor than its predecessor's.
A model that deliberately underperformed would probably not be detected. As OpenAI's chief scientist put it: progress in intelligence does not guarantee progress in alignment.
Our Take: Why Nobody Wants to Hit the Brakes
Why did OpenAI release Astra despite these risks? Simple: it is a race, and one with a calendar.
Fable 5.1 arrived on September 1, Astra two days later, with a clear message between the lines: whatever Fable can do, we can do too, and much more.
That also explains why Greg Brockman talks about AGI. OpenAI wants to be first to claim the territory.
Shortly after launch, reports said Anthropic's IPO, potentially the largest ever at a valuation of up to $2 trillion, had slipped to mid-October at the earliest.
The official reasons are financing and regulation, not Astra. There is also a rumor that Anthropic will release something big before the IPO. That would not surprise us.
Someone hitting the brakes would.
As we celebrate and use these new capabilities, the dystopian scenarios at ai-2027.com are becoming more real piece by piece.
We have to hold both ideas at once without letting one erase the other. It is uncomfortable, but it is the only honest position we can see right now.
We will keep watching.
That is it for today. This turned into a fairly dense issue.
If you made it this far: which task will you delegate this week?
Replies go directly to us as always. We read them.
Reto & Fabian from AInauten
