Howdy wizards,

Here’s what’s brewing in AI.

I tested Fable 5 this weekend to see if it could help me reduce some of the operating expenses in my business.

I did exactly what I was talking about on Friday: had Opus 4.8 plan everything, then had Fable 5 implement and test. It saved me around $200/yr in expenses in 30 minutes.

I actually had two challenges I wanted to solve, both are about reducing my costs for running contextwindows.ai. I gave each of the challenges below to Claude Opus 4.8:

I've explained these infrastructure tools (VPS, Vercel, Supabase) for building with AI in more detail in my post last week about the 8 AI tools I actually use.

Claude managed to cleanly map out a plan for both of these challenges.

However, I proceeded with only 1 of them (and Fable did a fabulous job).

I’ll highlight this because it’s the main thing I want to share:

In this day and age of AI being able to build and implement just about anything, knowing when not to proceed with something is just as important as knowing when to say “go”.

Also — if you have meetings, here's one thing totally worth a “go": a note taker that lets you pay full attentions to the humans on your screen instead of scribbling things down.

Which is where Granola, this week's sponsor, comes in:

IN PARTNERSHIP WITH GRANOLA

Turn meetings into momentum

Meetings aren’t the problem. It’s everything that comes after them.

The context switching, the cognitive load of remembering what you promised, the fear of something important slipping through…

Granola lets you take notes the way you always have. It works in the background, turning conversations into clear summaries, action items, and next steps.

With Granola, your meetings are no longer about scrambling to keep up. They lead somewhere.

Download Granola for free and use it in your next meeting →

❦

I started out happily hosting my product Context Windows on Vercel Pro at $20/mo.

The first month the bill was $25.

Then the month after about $30.

And it kept increasing month over month.

And now.. $50/mo.

Just look at all those line items. I’m gonna keep it real with you guys — I only understood that I was being charged $20/mo for Vercel Pro. I didn’t understand how any of those other charges happened.

So I put Claude Code with Opus 4.8 to the test. I attached my bill as a PDF.

“Audit my codebase and Vercel setup to understand all the points in the system that trigger these charges. I’m looking to reduce the bill back to $20/mo, without affecting any functionality of my product. Lay out a plan for how to implement measures that will drive the bill down”

The model thought for 14 minutes and came back with an understandable breakdown about how these credits were consumed, and a detailed plan on measures to reduce the bill.

After Opus 4.8 had done the legwork of auditing everything and pinpointing the places in the system to tweak, I had Fable 5 do the implementation. I also asked it to test everything after it was done, to make sure everything still works.

It came back after around 20 minutes…

There were basically lots of inefficiencies in my system, processes consuming credits with no functional upside, and Fable just… took care of it. Easily.

It even found a couple of related, latent bugs in my codebase and fixed them.

So.. $200+/yr saved in bills, done in well under an hour.

And keep in mind, the starting point for me was pure confusion about what the bill even represents, let alone where in my system those credits were consumed.

I’ve heard the cool devs on X talk about how it’s so simple to just self-host Supabase and pay $0/mo. I had thought about it for a while, but had the feeling it would be complex.

So I asked Claude Opus 4.8 to assess the feasibility of migrating to self-hosted Supabase.

It came back after 13 minutes with a clear recommendation: don’t do it.

Bottom line: migration is not a problem. It’s easy. But the consequences of migrating to self-hosted in my case, on the other hand, are a problem.

It creates a lot more management work for me, and also would need me to upgrade my VPS to a higher tier. It also means I can’t use the Supabase MCP, which I do all the time, and it makes my life a lot easier.

So I decided not to do it, and I consider that a success — now I don’t have to dream about migrating that thing anymore, and my focus can go back to actually being productive.

I go on this rodeo every week.

Get a new idea, plan it out, figure out if it’s worth doing.

Whether it’s possible to do is almost not a concern anymore. It almost always is.

But saying yes has consequences for cost, time and quality. And understanding those and how they relate to your why is the game.

Here are the 3 things to keep in mind before saying “yes, Claude, build it”:

Here’s how that stacked up for my two challenges:

Proceeding and not proceeding with something.

Both are great outcomes when they’re rooted in what you want.

So the recipe is this:

Point AI at the right problems, know why you’re doing it, and know where you can compromise and where you can’t.

❦

Enjoying the newsletter? Forward it to your friends 🤝

You are a delight.

Dario

Most companies pick AI use cases by brainstorming internally. 90% of those initiatives fail.

I’ve created Context Windows just so you can pick the winners.

🟦 Find high-performing use cases from 2,000+ companies at contextwindows.ai, or book a demo with me

Disclosure: To cover the cost of my email software and the time I spend writing this newsletter, I sometimes work with sponsors and may earn a commission if you buy something through a link in here. If you choose to click, subscribe, or buy through any of them, THANK YOU – it will make it possible for me to continue to do this.