Hello AInauts,

Welcome to the latest issue of your favorite newsletter!

This week we are going deep into the next generation of AI models.

Today we brought things you can apply right now:

Let's go.

When you type a prompt, you summarize. When you speak one, you explain. Wispr Flow captures your full reasoning — constraints, edge cases, examples, tone — and turns it into clean, structured text you paste into ChatGPT, Claude, or any AI tool. The difference shows up immediately. More context in, fewer follow-ups out.

89% of messages sent with zero edits. Used by teams at OpenAI, Vercel, and Clay. Try Wispr Flow free — works on Mac, Windows, and iPhone.

Start flowing free

We do not love writing about every new model update anymore.

Yes, every few weeks another newer, better AI model arrives. We are used to that by now.

This one could matter for a different reason: normal users may soon struggle to afford the strongest models.

First things first.

Anthropic has released Claude Fable 5. It is a new-generation model, one level above the previous top model. As usual, it is "state of the art" on almost every benchmark.

That is nice. We also expect it at this point.

Fable 5 is a version of the Mythos model, which has been making headlines for a while. Mythos finds critical software vulnerabilities that have remained hidden for years. Fable is a version with safety guardrails, which makes it usable for a broader audience.

After only a few hours, many people already called it the best coding model ever released.

Stripe says it compressed months of work into days.

Or you can throw it a screenshot of a web app, and it rebuilds the thing for you.

For months now, we have felt a real shift in our day-to-day work with the newer AI models.

What these systems can do is wild when you use them properly. The bigger surprise is how long they can work autonomously if you let them.

One example from our week:

We now define a goal with a precise description of the expected result, and the AI immediately starts analyzing, building dynamic workflows and launching 10 to 30 subagents.

Each one gets its own task. All of them run in parallel.

A few million tokens and an hour later, the result is there.

In our case, it was a full analysis of a new AINAUTEN approach, including competitor research, websites, structures, emails, financial considerations and more.

What already worked well with Opus works even better with the new Fable.

We keep having the same thought:

Most people out there have no real feeling for what is becoming possible week after week.

That much power costs money.

Fable 5 is about twice as expensive per token as the previous flagship model, Opus 4.8.

And the part people are arguing about: from June 23, it leaves the normal subscriptions again. Anyone who wants to use it after that will pay per token.

The internet immediately turned this into a full drama:

And yes, this may be the direction for a while.

These models need enormous compute. With 20-dollar subscriptions, Anthropic and others have likely been paying part of the bill themselves.

We are still more relaxed about it.

Anthropic says the reason is capacity. The model is so in demand right now that the servers cannot keep up. They want to bring it back into the subscription once that becomes possible.

We do believe one thing, though:

The very best models will probably become expensive. Really expensive.

Because the tasks that require hours of work and can now be handled reliably are also genuinely valuable.

If you have a Claude subscription, try Fable while you can.

Use it with the /goal command or give it a truly complex task.

You need to experience for yourself what these models can now do.

The era of the smart chatbot is ending. What matters now are agents that actively work for you.

And if the strongest models become expensive, one skill becomes more important:

knowing which model you need, when to save tokens, and how to organize your AI work efficiently.

We will be dealing with this for a long time.

AIApply is your AI Career Agent working 24/7 to find the best jobs online, tailor every application to your profile, and automatically apply on your behalf so you can spend less time job hunting and more time landing interviews.

Start landing interviews faster.

Something went viral again in the AI bubble this week.

Peter Steinberger basically said:

More than 8 million views.

The most accurate answer underneath was not an explanation. It was:

Boris means Boris Cherny, the person behind Claude Code.

Two of the smartest people in AI are saying: prompting is over, looping is the new prompting.

If you want the whole viral story, there is a good technical summary. But what matters to us is what this kind of trend means for people like you and us.

So here is the short AInauten version.

First, we learned to prompt well.

Then we learned to give AI the right context: files over tools, pulling knowledge out of chatbots and saving it into local folders, and so on.

Now a third layer is entering the conversation:

you stop prompting the AI yourself and build a loop that prompts it for you.

Some people already call this "loop engineering."

At its core, a loop is simple:

a small program that prompts the agent, reads the result, decides whether it is good enough and prompts again if it is not.

You are no longer the person typing inside the loop.

You write the loop.

Or, in the bluntest version:

In its simplest form, a loop is just a process that runs automatically at regular intervals.

You may already have one running in the background.

If you have set up a routine in Claude or Codex, or any scheduled task, you already use the simplest form of a loop.

So much for the mysticism.

Of course, the loops used by professionals can do more.

For us non-engineers, the useful question is not necessarily "how do I build loops?"

The better question is:

Loop-able work is work you can list clearly and where you can check whether a piece is finished.

More concretely:

loop-able tasks are often Claude or Codex skills that you currently trigger manually on a regular basis.

Some things are much harder to loop:

In those areas, your judgment is the value. A loop can give you confident drafts, but it cannot replace the decision.

The simplest form is really simple.

For example, in Claude Code you can write something like:

The structure is:

then the time interval, and then the task that should run.

In this example, the memory-audit skill runs every five minutes.

Right now, that is the most practical loop setup:

run well-defined skills on a recurring schedule.

The new skill is not necessarily "building loops."

It is recognizing which 20 percent of your work is loop-able and leaving the rest alone.

For most people, prompting is not dead. Looping does not replace it.

But the best of the best, especially coders, are already using this framework at a different level.

The /goal command is where this gets interesting too. It aligns the model around a target, and a goal often contains several loops.

For many people, this is still not the most relevant layer. It can get complex fast.

Tomorrow's Deep Dive goes deeper into this topic:

Ideal if you want to go deeper.

We will stay on it.

We have now talked about the new generation of models and also established that prompting is not dead.

So let us share a few prompting tips from Anthropic for this new generation.

You can apply them immediately.

The biggest shift:

the new models follow short, clear instructions better than ever.

That means the huge, nested mega-prompts many of us have been collecting for years can make results worse with the newer models.

We have our own prompt library too. Guilty.

Anthropic writes that overly detailed old prompts "can degrade output quality."

Short and clear beats long and complicated.

Instead of saying:

try:

The model can connect your task to the context instead of guessing what you actually want.

It sounds basic. It is one of the biggest levers.

Most people test new models with tiny tasks:

Then they conclude: okay, nice.

Wrong test.

Top models shine on hard, long, messy problems.

Give it the problem you have been avoiding for weeks.

That is where you will actually be surprised.

This is directly inspired by Anthropic's tips for Fable:

If you have enough information to start, then start.

That means less hand-holding and fewer clarification loops.

Tell the model to carry the task through end to end and deliver the result at the end, instead of asking you about every step.

We use exactly these approaches and see excellent results.

Our advice: just try it.

If you take one thing from this:

make your prompts shorter and give the model more context about the why.

That is it for today.

Hopefully not too technical. If it was, please let us know.

See you in the next issue,

Reto & Fabian from AInauten

---