Claude Fable 5's First Week: 8 Demos People Have Already Built

Last updated: June 2026

Claude Fable 5 went public on June 9. The launch coverage ran the usual route: benchmark tables, an IPO angle, a safety debate. All real, and none of it answers the question most people actually have about a new model, which is what it can do in practice.

So here is the other route: eight things people built with it in the first week, what each one actually demonstrates, and a link to every original post. Then the part most release coverage skips — none of this runs on hardware you own, and that has consequences worth understanding before the novelty wears off.

What Claude Fable 5 Is (and Isn't)

Anthropic released Claude Fable 5 on Tuesday, June 9, 2026. It is the first generally available model in the company's Mythos class — the tier Anthropic unveiled in April and deliberately kept away from the public, limiting access to vetted organizations through a program called Project Glasswing because of the model's offensive cybersecurity capability. Fable 5 is the same underlying model as the restricted Claude Mythos 5, with one structural difference: safety classifiers that block requests in specific high-risk domains, including cybersecurity, biology, and chemistry. When a request trips those classifiers, the system declines and falls back to Claude Opus 4.8.

Availability is in flux. As of publication, Fable 5 is live on Anthropic's API and consumption-based Enterprise plans, and included in Pro, Max, Team, and seat-based Enterprise subscriptions through June 22. On June 23, Anthropic says it will pull the model from those subscriptions and meter it with usage credits, with a stated plan to restore it as a standard subscription feature later. Anthropic reports a score of 80.3 percent on SWE-Bench Pro, ahead of every generally available model — a vendor-reported number, which is exactly why the demos below are the more interesting evidence.

A sourcing note before the list. These demos come from release-week posts on X, several of them amplified by Anthropic's own account, and every capability claim is the builder's own — ModemGuides has not reproduced them. Each demo below ends with a Load post from X button. Until you tap it, nothing from X loads on this page — no widget script, no trackers. Tap it and the post loads in place with X's Do Not Track flag set. If you would rather not load X content here at all, each card also carries a plain open on X link to the original.

The Demos at a Glance

All results in this table are builder-reported. Judge the footage at the original posts linked below.

What People Built in Week One

A soccer shot-speed coach from one prompt

The builder posting as AA (@measure_plan) handed Fable 5 a goal — typed, literally, as a goal: "help me shoot this soccer ball faster" — and got back a working analysis overlay for his own practice footage. The clip shows a live readout of shot speed (75 km/h on the measured kick), current ball speed, and the ball's distance to the goal line. The interesting part is not the sports science. It is that a vision-analysis tool fitted to one person's video existed minutes after being asked for, with no code written by the person asking.

Yosemite, navigable and to scale

Shlok Khemani (@shloked) asked Fable 5 to build a navigable version of Yosemite at real scale. According to his post, the model pulled satellite imagery and NASA elevation data, classified individual forest pixels to place roughly 266,000 procedural trees, and wrote custom water shaders for all six of the valley's famous waterfalls, positioned on the correct cliff brinks. Whether every figure holds up under scrutiny, the workflow it describes — source real geodata, classify it, generate assets, render the result — is a long, multi-step agentic chain. That chain completing at all is the demonstration.

The Boeing 747 test

This one has history. Victor Mustar, product lead at Hugging Face, has spent months running an informal test on frontier models: build a detailed Boeing 747 in Three.js using only the library's geometric primitives, no downloaded assets. As he framed it when he introduced the test, it is "more about spatial understanding than library knowledge," and through late 2025 he reported that no model got it right. On June 9 he posted Fable 5's attempt and called the result "AGI-level." Discount the superlative — this is an enthusiast's bar, not a controlled benchmark. But it is a bar this specific tester watched every previous frontier model miss.

A wooden house under deliberate constraints

Angel (@Angaisb_) stress-tests models on structured building tasks using a deliberately restrictive system prompt that, in their words, "removes creativity but makes buildings more functional" — a constraint harness, not a creativity showcase. Run at maximum reasoning effort against that harness, Fable 5 produced a wooden house that met the bar, and Angel called the results strong enough to keep testing against earlier prompt versions. The demonstration here is instruction-following under constraint: holding a strict specification across a long build instead of drifting toward whatever looks dramatic.

A soft-body rig you can drag around

Jason Kneen (@jasonkneen) built a character rig with soft-body physics that runs in the browser: drag the control points and the body responds with believable movement. What earns this one a slot is the builder's own honesty about the gaps — weight and gravity still need work, and the limbs read as what he calls "paper legs." Self-reported failure modes are rare during a launch week, and a demo whose author lists its problems is more informative than ten that don't.

A goldfish you can feed

Midori Tatsuta (@midori_tatsuta) posted, in Japanese, an app made with Fable 5 alone: a browser goldfish you can feed, with a public link so anyone can try it. It is deliberately trivial, and the triviality is the point. Software this small was never worth a developer's time, so it simply did not get made. When a complete, shippable toy costs one conversation, an entire category of disposable, single-purpose software starts existing.

A drowned city in one sentence of intent

Ethan Mollick, the Wharton professor whose AI experiments are widely followed, asked Fable 5 for a shader that could run in twigl.app — a niche browser playground for live-coding GLSL with its own compact format. The prompt asked for "an infinite city of neo-gothic towers partially drowned in a stormy ocean" with large waves. The result is fully procedural — no textures, no models — and when he wanted more, the entire iteration loop was telling the model to make it better. Knowing twigl's quirks at all is the quiet demonstration; most humans who write shaders don't.

Liquid-glass UI in pure CSS — on Mythos 5

Ann Nguyen (@ann_nnng) posted glass-refraction interface effects — the liquid, light-bending look — built in CSS alone, no WebGL. One clarification before you file this under Fable: she credits Mythos. Claude Mythos 5 is the same underlying model as Fable 5 with the safety classifiers removed, and it is available only to organizations approved through Project Glasswing. So this demo doubles as a reminder that the most capable version of this model is one most people cannot access. The gap between the tiers is policy, not weights.

The Pattern: Prompt-to-App

Strip the subject matter away and the eight demos share a shape. Each is a single session producing a complete, browser-deliverable artifact — no development environment, no dependency management, no deployment step. Iteration happens in plain language rather than code review. And the builders span the range from professional developers to people who simply described what they wanted.

That shape is the actual release-week news. Output that used to require a developer and a toolchain now arrives as a conversation, and the floor for who can produce working software dropped accordingly.

Two caveats keep this honest. These are highlights selected by motivated builders during a launch window, several boosted by the vendor's own account — they establish a ceiling, not an average. And every claim above is self-reported; the links exist so you can judge the footage yourself.

The Local-First Reality Check

Here is the sentence this site owes you: nothing you can download and run on your own hardware does what these demos show, end to end, today. Fable 5's weights are closed and its inference happens on Anthropic's servers. The best open-weight models running on a capable home rig — a used RTX 3090 build, a 64GB AMD mini PC — are genuinely useful for coding help, document work, and private chat, and our guide to local AI hardware covers exactly what that money buys. It does not buy this.

What the closed frontier costs you is control. Every prompt, every frame of that soccer video, every document routed through these demos transits someone else's infrastructure — the same exposure model we examined when Claude Code's own source code leaked through npm. And the terms move under you: Fable 5 launched inside paid subscriptions on June 9 and is scheduled to leave them for usage credits on June 23. Anthropic says the change is temporary. Either way, that is what renting capability looks like — the landlord sets the terms, and the lease changed in week two.

So the honest framing is a trade, not a contest. If you want the frontier, today you rent it. If you want ownership — your data on your network, your model unchanged until you change it — the local stack is real, improving, and cheaper than it looks; start with our mini PC roundup for local AI. Know which trade you are making, and make it on purpose.

Frequently Asked Questions

What is Claude Fable 5?

Claude Fable 5 is Anthropic's most capable generally available AI model, released June 9, 2026. It is the first public model in the company's Mythos class and ships with safety classifiers that block requests in high-risk domains such as cybersecurity, biology, and chemistry.

What is the difference between Claude Fable 5 and Claude Mythos 5?

They share the same underlying model. Fable 5 adds safety classifiers and is broadly available; when a request trips a classifier, the system declines and falls back to Claude Opus 4.8. Mythos 5 omits those classifiers and is restricted to organizations approved through Anthropic's Project Glasswing, primarily cyber defenders and critical-infrastructure operators.

Can I run Claude Fable 5 locally or download its weights?

No. The weights are closed and inference runs only on Anthropic's infrastructure. The nearest local path is running open-weight models through a tool like Ollama on capable hardware — our local AI hardware guide covers what different budgets realistically buy. A home setup will not match Fable 5's ceiling, but it is fully yours.

Are these demos verified?

They are builder-reported. ModemGuides has not reproduced them, and several were amplified by Anthropic's official account, which is worth weighing. Every section above links the original post so you can watch the footage and judge the claims directly.

How can I try Claude Fable 5?

As of publication, Fable 5 is available through Anthropic's API — listed at launch at $10 per million input tokens (Anthropic, June 2026) — and on consumption-based Enterprise plans. It is included in Pro, Max, Team, and seat-based Enterprise subscriptions through June 22, 2026, after which Anthropic says access from those plans will require usage credits until subscription access is restored. Check Anthropic's current pricing before committing.

What happens when Fable 5 refuses a request?

In domains the safety classifiers flag — cybersecurity, biology, and chemistry among them — Fable 5 declines and the request falls back to Claude Opus 4.8, meaning the answer you receive comes from a different, less capable model. For developers, Anthropic's documentation notes this changes refusal handling, retry logic, and billing for API integrations.