Make the world's smartest AI train its replacement

Edition 192 - The operator move that makes Fable 5 worth the price tag. Once.

Author

The AMP Team
July 07, 2026

Here’s what we’re reading and thinking about this week:

Claude's Fable 5 has been the talk of the town the last week. For example, Stripe says it compressed months of engineering into days.

It's also the most expensive AI model on the market. It will eat up your tokens quickly or surprise you with a hefty end of month bill.

But when we handed it our newsletter playbook, the output was indistinguishable from when we run it on cheaper models.

Our first reaction: Fable isn't built for this kind of work.

So what kind of work is "worth" the expensive price tag? And better: how do you make sure you only pay it once?

Quick catch-up if you missed the drama: Fable 5 is Anthropic's newest model, and the smartest AI ever released to the public. It launched in early June, got pulled by export controls, and came back worldwide on July 1.

Here's the thing about the success stories: almost all of them come from engineering teams. That's not because AI only likes code. It's a clue about what this model needs to do its best work. And it's something you can give it.

The miracles happen where the outcome is checkable

Fable's edge shows up when two things are true: success has a yes-or-no answer, and getting there is genuinely hard.

Stripe's project is the perfect example. They described an issue that was fifty million lines of code, a change that would take a team two months by hand. But the outcome was checkable: the code works or it doesn’t.

So the AI model could grind for hours, check its own work, and keep going without a human approving every step.

Hard path, checkable outcome. That's the Fable sweet spot.

Your work isn't on the other side of coding... It's on a spectrum.

Most knowledge work gets judged differently. There's no test that tells you a campaign brief is correct. Oftentimes, quality lives in how the work was done, and "good" is a judgment call that sits in someone's head.

But your work isn't all one thing. Data cleanup, number-checking, and analysis sit near the checkable end. Fable scored 10 points higher than the last model on a hard financial analysis test, and that's knowledge work, not code.

When we tested a “checkable” task - our webinar performance report - Fable did great.

But again, our newsletter had really no improvement.

So why did the world's smartest model change nothing?

What our newsletter test actually told us

Our newsletter playbook is specific. Inputs, steps, voice rules, quality standards, all written down. When we ran it on Fable, the result looked like it always does.

That's not Fable failing. That's the playbook working.

When the work is fully written down, any good model can run it, and the expensive one adds nothing. Which is exactly what you want: own the playbook, rent the tech.

So Fable's price tag is significantly more meaningful on work you haven’t playbook’d yet.

That changes the question. "Should we use Fable" isn't about the model. It's about your work: what do you have that's hard, checkable, and still stuck on your desk?

The bottleneck isn't how smart the model is

Our take: the bottleneck is whether anyone wrote down what good looks like.

Engineers aren't winning because their work is special. They're winning because they spent time writing down their standards in a way a computer can check. They call it a test suite.

Every AI "failure" we've had comes back to the same thing: we hadn't decided what we wanted clearly enough to direct it. A smarter model can't fix that.

Writing it down can.

How to make your work Fable-shaped

The operator move is to pull the standards out of your head and into the playbook.

The playbook is the knowledge worker's test suite. And once it exists, you get to make a real decision instead of a vibes-based one.

Everyday runs go to the cheaper model, because a well-written playbook runs the same everywhere.

Fable gets the big swings: the quarter of data you never had time to dig through, the review of every process doc you own, the analysis that's heavy, checkable, and still stuck.

That's what an AI operator actually uses Fable for. The magic isn't in the model. It's in handing it work where it can tell whether it succeeded.

Make Fable train its replacement

Now for the promise we opened with: paying the Fable price only once.

AI labs have a move called “distillation”.

They use their biggest, most expensive models to teach smaller, cheaper ones. You can run the operator version of the same play.

When Fable does something great for you, don't just take the output. Have it write down how it got there: the inputs it used, the steps it followed, the checks it ran. That document is a playbook draft. Correct it until it's right, and next time the work runs on a cheaper model.

You pay for the genius once. The playbook keeps it.

Do this this week

  1. Test Fable out on something “checkable” that you haven’t playbook’d yet. The harder the better.

  2. If it does a good job - have it write a playbook from it. Then run that in a cheaper AI model and see how it goes!

Hit reply and tell us how it goes!

UPCOMING EVENTS

Have you started playbooking yet? Join our free masterclass next week!

Learn how to turn your most repetitive processes into AI-powered workflows with our playbooking method.

LINKS

For your reading list 📚

That's all!

We'll see you again soon. Thoughts, feedback and questions are much appreciated - respond here or shoot us a note at [email protected]

Cheers,

🪄 The AMP Team (formerly: the AI Exchange Team)