What People Built with Claude Fable 5 in Its First 72 Hours

TL;DR: Three days after Claude Fable 5 launched on June 9, 2026, Anthropic's official account published a showcase of what the community had already built—and it racked up 4.2 million views in hours. The projects skew heavily toward creative and physics-heavy coding: procedurally generated shaders, fluid simulations, real-time physics rigs, and small playable apps shipped end-to-end. The pattern across all of them: builders are throwing deliberately hard, open-ended challenges at Fable 5 and finding that it clears them on the first or second attempt.

The Showcase

On June 12 at 9:00 PM, the @claudeai account posted: "Claude Fable 5 has been out for a couple of days. Some projects people have already built with it." What followed was eight quoted projects from the model's first 72 hours. Here's each one, and why it's more interesting than it looks.

1. Soccer-ball physics, optimized by /goal

The first project came from @measure_plan, who gave Fable 5 a single instruction through Claude Code's goal mode: "help me shoot this soccer ball faster."

What makes this notable isn't the soccer ball—it's the workflow. The /goal command defines a verifiable end state and lets the agent iterate autonomously until it's met. Instead of prompting turn by turn ("try increasing the force coefficient"), the builder stated an outcome and let Fable 5 run the experiment loop itself: adjust parameters, simulate, measure, repeat. It's a small, legible example of the loop engineering pattern applied to a physics toy.

2. Ethan Mollick's drowned gothic city

Wharton professor Ethan Mollick—one of the most-watched AI experimenters—gave Fable 5 this prompt for twigl.app, a live-coding shader environment:

"Create a visually interesting shader that can run in twigl.app. Make it like an infinite city of neo-gothic towers partially drowned in a stormy ocean with large waves."

Then the follow-up that's become his signature test: "Make it better."

The result—an entirely procedurally generated, infinite cityscape with animated storm waves—matters because shader programming is brutally unforgiving territory for LLMs. There's no framework to lean on, no library calls to memorize; it's raw math (signed distance fields, ray marching, noise functions) that either renders something beautiful or renders garbage. And "make it better" is the purest test of aesthetic judgment a model can face: no specification, just taste.

3. Fluid ink that melts together

Japanese indie developer @hayashimon1 set out to find Fable 5's breaking point: "I deliberately threw a tough challenge at Claude Fable 5 to see how far it could express itself. An effect where ink melts together like a fluid."

Their verdict, translated: "I thought this might be tough and went to test its limits, but it just turned into a proper form without any issue."

Fluid simulation is another no-shortcuts domain—convincing ink diffusion means implementing something like Navier-Stokes-adjacent advection or a clever approximation, then tuning it until it feels liquid. A builder explicitly trying to make the model fail, and reporting that it didn't, is a stronger signal than any staged demo.

4. The Boeing 747 benchmark

Victor Mustar (Hugging Face's head of product) declared that "Fable has done an AGI-level job on the Boeing 747 benchmark... it's almost scary."

The "Boeing 747 benchmark" is one of those informal community evals—like the pelican-on-a-bicycle SVG test—where everyone asks every new model to build the same absurdly complex object and compares results. They persist because they work: a 747 has landing gear, engine nacelles, a distinctive hump, and hundreds of spatial relationships that can't be faked with memorized code. We covered why these vibes-based evals coexist with formal ones in our complete guide to AI benchmarks.

5. A goldfish app you can play right now

From @midori_tatsuta in Japan: a complete goldfish-feeding app, built entirely with Fable 5 and deployed to Netlify for anyone to play.

It's the humblest project on the list and maybe the most representative. Not a tech demo—a finished, shipped, playable thing made by one person in a day or two. The distance from "idea" to "URL you can send your friends" is the metric that matters for most people, and it keeps shrinking. This is the same energy as the full-stack websites people build with Claude, compressed further.

6. Jason Kneen's responsive physics body

Developer Jason Kneen showed a rig where dragging control points around makes a simulated body respond with realistic movement—while candidly noting what's unfinished: "Still work to do on weight, gravity and the 'paper legs.'"

The honesty is the point. This is what real iterative development with a frontier model looks like: impressive core behavior, visible rough edges, a builder who knows exactly which ones. It's a healthier signal than polished launch-day cherry-picking.

7. "A wooden house" at max reasoning effort

Angel (@Angaisb_), known for running the same structured building tests against every new model, put Fable 5 (max) through the classic: "A wooden house."

Two details make this one interesting. First, it was run at max reasoning effort—the setting that trades latency for deeper deliberation, which we break down in our effort parameter and model selection guide. Second, it used Angel's deliberately constrained "safe" system prompt, which removes creativity and pushes the model toward functional architecture. Passing a test that's rigged against flashiness—and still impressing the tester—says more about structural reasoning than any freeform demo. Their verdict: "I'll keep testing it with other past system prompts, but so far so good."

It's the same capability surface we explored in Fable 5's Minecraft 3D worldbuilding: translating a one-line brief into a coherent, buildable structure.

8. Yosemite, navigable and to scale

The most ambitious project in the showcase came from Shlok Khemani (@shloked), who asked Fable 5 to build a navigable version of Yosemite—to scale. The model's pipeline, per his thread:

• Pulled satellite imagery and real NASA elevation data
• Classified individual forest pixels and generated ~266,000 procedural trees
• Wrote custom water shaders for all six famous waterfalls, placed accurately on their cliff brinks

This is a different kind of demo from the shaders and toys above. It's agentic: the model didn't just generate code from imagination—it sourced real geospatial datasets, processed them, and assembled a verifiable artifact. Anyone who has hiked Yosemite can check whether Bridalveil Fall is on the right cliff. That's ground truth as a test harness.

The Pattern: Hard, Verifiable, Visual

Look at what unites these eight projects:

Every one of them is self-verifying. The output is visual, interactive, or measurable—you don't need a leaderboard to judge it. That's why the community gravitates to these tests within hours of any launch, and why they're harder to game than static benchmarks (a tension we explored in specification gaming and Goodhart's law).

It also rhymes with the launch itself: Anthropic let Fable 5 edit its own launch video—the same "show, don't claim" instinct.

What This Signals About Fable 5

Three days is too early for rigorous conclusions, but the early returns line up with what we found in our top 10 Fable 5 use cases analysis:

Creative-technical work is the standout. The model's edge shows most in domains that mix math, code, and aesthetic judgment—shaders, simulations, 3D worldbuilding. That tracks with a Mythos-class model's deeper reasoning being applied to problems where every step compounds.

Autonomy is becoming the default workflow. Three of the eight projects ran through goal-directed, iterative, or data-sourcing agentic loops rather than turn-by-turn prompting—the soccer-ball /goal run, Mollick's "make it better" iteration, and the Yosemite pipeline. If you're building this way, our guide to Fable 5 loop design, self-correction, and memory covers the patterns.

Non-engineers are shipping. The goldfish app and the showcase's Japanese-language entries underline that the first-wave builders aren't just Silicon Valley engineers. The interface is natural language; the output is deployed software.

If you want to try it yourself, start with Is Fable 5 available in Claude Code? (yes, with setup details) and the launch announcement breakdown for pricing and model-tier context.

The Competitive Backdrop

The showcase landed in a busy news week: a day earlier, Google Research announced Gemini-SQL2, its state-of-the-art text-to-SQL capability. The contrast in launch styles is instructive—Google led with a benchmark number and no public access; Anthropic led with what strangers built in 72 hours using a model anyone can access today. Both are valid evidence. But only one of them comes with a goldfish app you can play.

Where to Go Next

• Claude Fable 5 and Mythos 5: Launch Breakdown — pricing, safeguards, and what Mythos-class means
• Fable 5: Top 10 Use Cases — where the model excels in practice
• Is Fable 5 Available in Claude Code? — setup and access guide
• Fable 5 Edited Its Own Launch Video — the launch's other "show, don't claim" moment
• What Is Loop Engineering? — the workflow behind the /goal soccer-ball project