The AI Filter: A New Model Took the Top Spot This Week. Here Is What That Changes for You.

OpenAI shipped a new flagship model this week, and within two days one newsletter told me it "feels like AGI," another told me the AI guru bubble had popped, and a third told me a chip executive says AGI is already here. Same week, same inbox. Here is what survived a second look.

This is The AI Filter, the weekly read where I go through the stack of AI newsletters so you do not have to. I cut the hype, keep the signal, and give you the reasoning so you can judge for yourself. This week there was a real event under the noise, and it is worth ten minutes of your attention.

The event is GPT-6 Astra, OpenAI's new top model. The Batch, Andrew Ng's weekly newsletter and the most careful source in my stack, laid out the independent test results. Astra leads some leaderboards outright. On the broadest composite index it landed in a virtual tie with Claude Fable 5.1, Anthropic's model, and the two of them are trading first and second place depending on which version of the test you read.

Now the filter. The subject lines that said "feels like AGI" and "$12,000 in 30 days" came from people selling courses and communities. The Batch's numbers came with costs and minutes per task attached. When a source shows you the cost of the result, you are reading instruction. When it shows you the result and a countdown, you are reading a sales page.

Three things this week were worth keeping.

The first is that the top spot changes monthly now, and that is a reason to relax, not to scramble. Two years ago a new model launch meant a real gap between the leader and everyone else. This week the top two are a coin flip on the composite scores. That means the model you pick matters less than the work you put it to. If you have built a good habit with one tool, you will not be left behind by keeping it. Chasing the leaderboard is a hobby. Doing your work with the tool in front of you is the job.

The second is a report Anthropic published, and it is the most useful thing I read all week. It is a plain accounting of what bad actors tried to do with Claude over eight months and how the company shut it down. Rocket guidance software. A surveillance system built for a foreign spy agency. Thousands of fake dating profiles. Competitors running the model through fraudulent accounts and selling the answers as their own. Read that list once and you understand something the hype never tells you: a tool that can do real work for you can do real work against you. The same week, I spent a Saturday cleaning adware off my own laptop that had come in through a browser extension I never thought about. The lesson is the same at both scales. Capability cuts both ways, and the boring hygiene, who has access, what did you install, what did you allow, is where most of the risk actually lives.

The third came from the top of The Batch itself, a letter about what the valuable skill is now. Ng's argument is that the people who get the most out of AI are not the ones typing the cleverest prompts. They are the ones who decide what to build, in what order, and how to know if it worked. He calls it shaping the build. I spent twenty-six years calling it something else. Before you build a course you decide what the learner has to be able to do, then you design the shortest path to it, then you check whether it worked. That method has a name in my world, and it has not changed because the tool changed. The most valuable AI skill in 2026 is the one instructional designers have been using since before the internet.

Notice what none of the three are. None of them is a tool to buy this week. None of them expires Friday.

The models are now close enough that which one you use matters less than what you do with it. Pick one, build the habit, mind the hygiene, and spend your attention on deciding what the work is.

If you have the expertise and you are ready to turn it into something you can teach, The Foundation is our free starter course. It is at https://rhynowerks.ai