What Can Fable 5.2 Build? Games, Excel and Editor Clones

What can Fable 5.2 actually build from a short prompt?

In the demos creators published after Anthropic released it, Fable 5.2 built a playable one-level Gears of War clone in Unreal Engine, a multiplayer Fall Guys-style game with 10 levels, a browser-based Excel clone with working formulas and pivot tables, and a desktop video editor modeled on DaVinci Resolve. It also made ray-traced physics simulations, motion graphics, and a 3D house in Blender. Some results missed: a tree in the middle of a driveway, an awkward animation, and a failed event recreation.

How much better is Fable 5.2 than the model before it?

Better, but not in a way that changes what an AI model is. One reviewer put it plainly: Fable 5.2 is a "non-insignificant bump" over Opus 5.5, Anthropic's previous top model, and not an "AGI is here" moment.

The numbers he cited back that up. On one agentic coding benchmark, Opus 5.5 scored 45.5 and Fable 5.2 scored 49.1. On another, the scores were 62.3 and 69.9. Both are gains of roughly 10%. He said gains on scientific research and business workflow tests were larger, around 15 to 20%.

Anthropic released a second model, Mythos 2, alongside it, but regular users only get Fable. Mythos is aimed at people with cybersecurity backgrounds.

Keep that scale in mind when you read the demos below. They show what a modestly better model can do when someone spends heavily on it.

What did Fable 5.2 build, and what did it cost?

These figures come from two creators who ran the demos. Costs are estimates at API pricing (the pay-per-token rate developers pay), not what a subscriber necessarily paid. Token counts for some projects were not given.

A separate set of smaller demos (a typography toy, GPU simulations, a fake desktop operating system) cost, by that reviewer's estimate, roughly 30 to 50 cents in tokens in total.

How good were the games?

The Gears of War clone was a single level you could play: take cover behind barriers, shoot, do the timed reload, get revived by an AI teammate. The creator said the lighting, textures and enemy AI worked well. He also said the animation for reaching over a barrier "looks kind of awkward." He called it fun enough that he wanted to keep playing and lowered the difficulty.

The prompt was three lines. It asked for the most accurate clone of the game using Unreal Engine, only one level, but matching the real game's graphics, gameplay and characters. He ran it in a high-effort coding mode and let it run.

The Fall Guys clone, Tumble Beans, was tested live with three people in multiplayer. It had multiple game modes, including a spinning log obstacle and a "Gate Crash" door mode, and the physics (bouncing, knockback) was a focus. The prompt was short: build a Fall Guys clone as a full parody, make the levels fun, complex and different, build 10 levels. He added multiplayer afterward. One player briefly wondered why right click jumped, which is the kind of small quirk you would expect.

Can it really build an Excel clone or a video editor?

According to the creator, yes, with caveats about who is judging.

The Excel clone runs in the browser. Cells hold real formulas you can click into, there is a sample sales tracker with fake order data, and the pivot table worked. The creator said he could not find a missing feature or flaw. That is one person's testing, and he planned to publish it so others could check. Treat "perfect" as a claim to verify.

The video editor, named Kestrel, is modeled on DaVinci Resolve. It runs locally on a Windows PC, not in a browser. It shows waveforms, splits clips, snaps clips together magnetically, supports keyframes, zoom and keyboard shortcuts. A teammate named Alex separately built his own vibecoded video editor and has been editing videos with it, which is a more meaningful test than a demo.

Both are clones of mature products. A working clone is not the same as matching years of polish, edge cases and support in the original.

Where did Fable 5.2 fall short?

The failures in these demos were small or specific, not catastrophic, but they matter:

• The tree in the driveway. In the Blender house, the model placed a tree in the middle of the driveway. The pool, glass, reflections and rooms looked good, but a human would have caught the tree immediately.
• The Roadster recreation. Asked to rebuild a Roadster event from pictures and videos, the result was, in the creator's words, "the first bad one I've seen."
• The awkward animation in Gears of War when reaching over cover.
• Early-stage textures. The Blender house looked "wonky" until textures were applied, a reminder that outputs can need a pass or two.

The pattern: it did well on well-documented things like game genres, spreadsheets and physics. It did worse on recreating a specific real-world event from media. Spatial common sense (don't put a tree in a driveway) slipped through.

What kind of prompts worked?

The creators favored short, loosely specified prompts, and said so. One said he likes "under specifying" prompts. The prompts shown:

Build a Fall Guys clone. Full parody. Make sure the levels are fun,
complex, and different. Build 10 levels.

Build a 3D render of a beautiful house in Blender. Make it so it
feels like a real estate developer's dream.

The water balloon prompt was longer and more technical. It asked for a ray tracing simulation of a bullet piercing a water balloon, with freeze-frame and free camera, sliders for bullet speed, balloon size, lighting and gravity, and one hard rule: no Three.js or external web libraries, code everything from scratch. The result was a single HTML file using WebGL2 (a browser graphics standard).

Three habits stand out. State the target ("clone of X"), set one or two hard constraints (engine, no libraries, a weight limit), and ask for adjustable controls such as sliders. For the 3D-printed tower, constraints included printer model, filament and a weight cap of under 10 grams.

What does this mean for you?

If you work with spreadsheets, reports or internal tools: the Excel clone shows that a working spreadsheet-style app is now within reach of one model and one long session. Most people don't need a full Excel clone. They need a small tracker that fits their process.

If you do design, video or marketing work: motion graphics explainers in 45 minutes for about $25 change the economics of internal training videos and explainers. Review the output carefully. Misses like the driveway tree are the kind of thing a reviewer must catch.

If you manage budgets: the headline projects cost five figures at API prices. The small ones cost tens of dollars. Ask what a build costs before approving it, and start small.

What you can do now: try a small, bounded request, such as a calculator, a simple dashboard or a one-page explainer. Write a short prompt, add one or two hard constraints, and check the result for common-sense errors. Anthropic also showed a free usage reset notification to users when they first chatted with Fable 5.2, so check your account before paying.

Can you build your own tracker instead of paying for another subscription?

If the Excel clone made you think about the tools you rent, the practical version is smaller. Describe the tracker, request form or dashboard your team uses, and have it built for you. Remy is an AI that builds and deploys full-stack business apps from a plain-English description. It produces the app with a database, sign-in and a live URL, and you own the code. Something like an approvals tracker or a simple sales dashboard is a more realistic first project than a full spreadsheet clone, and it replaces a spreadsheet or a subscription you may be juggling today.