I Put Fable 5.2 to Work: 5 Real-World Field Tests and What I Learned

Fable 5.2 is the best AI model I've ever used, by far.

After spending days pushing it to its limits, I'm convinced we're entering a completely different era of software development. 

I gave Fable a three-line prompt and got a playable Gears of War clone with incredible graphics, enemy AI and real combat mechanics. I asked for Microsoft Excel and got the most feature-complete spreadsheet clone I've ever built. My team created everything from professional video editors to multiplayer games, custom Android utilities and complex 3D environments.

These were working applications, not mockups, and some are already being used in our actual production workflows.

What impressed me most: Fable takes massive, complex, underspecified assignments and figures out how to execute them. The level of autonomy and the quality of the results are unlike anything I've experienced. 

What disappointed me: The resource consumption is insane. My Gears of War project burned through 8.5 billion tokens, equivalent to roughly $15,000 in API costs. Even the best results weren't flawless.

Who should care: Developers, game creators, designers and anyone who's ever had a software idea but lacked the technical skills to build it.

Fable 5.2 is absolutely a meaningful leap over previous models. I've built versions of some of these projects before, and nothing came close to this level of completeness.

When I say “software is solved,” I don't mean every application is perfect. But technical ability is no longer any kind of barrier. If you can imagine it — and you have the tokens — you have a real shot at building it.

The biggest leap in Fable 5.2 is that it can take a ridiculously ambitious assignment and keep working until it has something functional to show you.

I've tested plenty of models that can generate impressive prototypes, but Fable 5.2 goes much further. Give it a short, surprisingly vague prompt and it can plan the architecture, source assets, build complex systems and iterate on the result with remarkably little hand-holding.

For my Gears of War clone, I gave it three lines of instructions, switched on Ultra Code mode and a few days later, I had a playable 3D shooter with cover mechanics, enemy AI and timed reloading.

That's a different level of autonomy.

The other major improvement is completeness. Previous generations could produce decent games and applications, but Fable 5.2 delivers far more sophisticated physics, graphics, interfaces and interconnected features. My latest Fall Guys clone, for example, is dramatically better than the versions I've built with earlier models.

Of course there's a catch: Fable's impressive results come at an enormous computational cost. My Gears of War project consumed 8.5 billion tokens, equivalent to roughly $15,000 at estimated API rates. 

It’s worth noting that other projects were surprisingly affordable, including a complex motion-graphics explainer for approximately $25. The breakthrough isn't that Fable makes every project cheap or effortless but that projects previously beyond the reach of an individual creator are now technically achievable.

Field Test 1: Personal Software
I asked, “Can Fable replace software I’m already paying for?”
Build stats — Excel clone: ~ 50 hours | ~8 billion cached-input tokens + 15 million output tokens | Estimated API cost: $11,700

I've been testing AI coding models for a long time, and one of my favorite challenges is asking them to recreate popular software.

Usually, you get something that looks pretty good until you start clicking around. The interface is there, but half the buttons don't work, the important features are missing and you definitely wouldn't replace the original software with it.

Well, Fable 5.2 just changed that.

I tested clones of Excel, DaVinci Resolve and Photoshop, and my producer, Alex McCraw, went even further and built his own versions of Adobe Photoshop and Premiere. 

The Assignment

Let's start with Excel.

I wanted to see whether Fable could build a fully functional spreadsheet application in the browser, not just a grid of cells, but something with the features people actually use: formulas, multiple sheets, data management and pivot tables. I’ve made an Excel clone a few times before, and none of them came close to the quality and feature-completeness of this version.

My prompt was ridiculously simple:

Build a full Excel clone in the browser. It should have full feature parity. Read the spec docs to figure out all the features.

That's it.

Meanwhile, Alex took on a different challenge. He wanted to recreate Photoshop and Premiere, but with a twist: instead of making one-to-one copies, he wanted the apps customized to his own workflow.

The Excel clone was very good.

It took 50 hours and 8 billion tokens, but Fable generated a browser-based spreadsheet with the familiar Excel ribbon and tabs: Home, Insert, Draw, Page Layout, Formulas, Data, Review, View, Automate and Help.

I could open an example sales tracker, work with the data, and interact with a pivot table.

I could also create another sheet, enter numbers and use formulas. I didn’t verify every feature in Microsoft Excel, but my clone behaved exactly as I expect a spreadsheet to.

Alex's experiments were even more interesting because they weren't just demos with fake data. He built a Photoshop alternative called Safelight.

Alex has been using Photoshop for years, so he knew which features he wanted and which he didn’t. He had Fable configure the controls, keyboard shortcuts and editing tools to the way he already works.

In his test, he opened a 4K document and created a YouTube thumbnail. He imported and resized images, manipulated layers, added text, used brushes and fill tools and worked with gradients.

And it wasn't some one-off experiment — Alex has already used Safelight to make several thumbnails for my channel.

Then he built a Premiere-style video editor.

This one has Claude integrated directly into the application. Instead of using an external AI agent to operate a traditional editor, Alex can describe the changes he wants and have the AI work within the software itself.

He's been using that application to edit videos, too. After a decade of succumbing to Adobe’s ruthless subscription model, Fable 5.2 has allowed Alex to finally cancel his subscription.

I also tested a DaVinci Resolve-style video editor called Kestrel. It ran locally on a Windows PC rather than in the browser, and the demonstration included a timeline, waveforms, clip splitting, magnetic snapping, keyboard shortcuts, zoom controls and keyframes. It worked perfectly.

For decades, we've bought software platforms built for millions of people. Product designers decide which features to include, how the interface should work and what we're allowed to customize.

But what happens when AI can build software for one person? As Alex points out, only a fraction of Photoshop users need every feature. 

Does this mean you should cancel every software subscription tomorrow? Obviously not. There's a massive difference between recreating the functionality you personally need and replacing every feature, integration and reliability guarantee of a mature professional application.

But I've already started using software I built with AI instead of the commercial alternative, and that’s changed how I think about software.

I really believe we're heading toward an era of “software for one”: You need an application, you describe it and AI builds it around your workflow.