10 Impressive Things People Built with Claude Fable 5 in Its First 48 Hours

On June 9, 2026, Anthropic released Claude Fable 5 — the first generally available model from its most capable "Mythos-class" line, and by several early accounts the strongest coding and agentic model the public has had its hands on. It shipped the same day inside GitHub Copilot and on AWS.

Two days in, the demos and customer reports are genuinely striking: hours-long autonomous runs, simulations derived from first principles, and codebase migrations that used to take a team months. Below are ten of the most impressive, each with a source. We have flagged which are vendor demos versus independent reports — and called out one viral showcase that may not be real at all. We build production MVPs for a living (our Takeoff program), so we read these less as hype and more as a preview of what the next year of product work looks like.

1. Stripe migrated a 50-million-line codebase in a single day

The headline result in Anthropic's launch announcement: Stripe used Fable 5 to run a codebase-wide migration across roughly 50 million lines of Ruby in one day — work the company estimated would have taken a team more than two months by hand. For anyone who has lived through a framework migration, that is the line that stops you cold.

Why it matters: the unglamorous, high-risk work — migrations, upgrades, dependency sweeps — is exactly where a model that can hold an entire codebase in context and verify each step earns its keep.

2. A week-long, mostly autonomous genomics run

Anthropic reports that its sibling model, Mythos 5, ran for over a week of largely autonomous work assembling single-cell data for millions of cells across 138 animal species — and built a custom machine-learning model about 100 times smaller than a published Science paper's model while outperforming it (source).

Why it matters: "runs for a week without losing the thread" is a different category of tool than "autocompletes a function."

3. It finished Pokémon with vision alone

Earlier Claude models needed an elaborate helper harness to stumble through Pokémon. Fable 5 completed Pokémon FireRed using game screenshots and a minimal harness (Anthropic). On the deck-builder Slay the Spire, giving it persistent file-based memory roughly tripled how often it reached the final act versus the previous Opus model.

Why it matters: vision plus memory is the combination that makes agents useful on real, messy, stateful interfaces — not just clean APIs.

4. A solar system simulation that predicts eclipses

In an Anthropic demo, Fable 5 built a solar-system simulation that derives the planets' orbital motion from physics first principles and uses it to predict solar eclipses (source) — not hard-coded values, but a model reasoning from the laws up.

5. A browser CAD tool that designs 3D-printable parts — and built itself

Anthropic showed Fable 5 designing complete, 3D-printable models inside a browser-based CAD editor — where the editor, including its built-in AI copilot, was also written by Fable 5 (source). A tool that builds the tool, then uses it.

6. Drug design accelerated roughly tenfold

Anthropic's internal protein-design experts say Mythos 5 sped up parts of the drug-design process by about ten times, matching or beating skilled human operators and generating strong candidates for 9 of 14 protein targets it was given. In blind comparisons, scientists preferred its molecular-biology hypotheses around 80% of the time over the previous Opus-class models (source).

7. Twelve-hour coding runs and one-shot full apps

Reviewers report Fable 5 sustaining single tasks for up to twelve hours against multi-page specifications (Digital Today). Builders on vibe-coding platforms told TechCrunch it is markedly better at "one-shotting full apps" with strong tool-calling, and developer Simon Willison shared early hands-on impressions the same day.

Why it matters: the practical unlock for founders is fewer, sharper prompts — describe the whole thing once, let it run, then review.

8. Autonomous factory-building in Factorio

Fable 5 played Factorio — the notoriously deep factory-automation game — on its own, planning and building a working automated factory without step-by-step guidance (Anthropic). It is a stress test for long-horizon planning, and a fun one.

9. Topping the benchmarks professionals actually use

Beyond demos, named teams reported real benchmark wins: coding lab Cognition logged the top score on its FrontierCode evaluation of production-grade coding tasks; analytics company Hex said Fable was the first model to clear its bar on complex, long-running analytical work; and trading firm IMC and Hebbia's finance benchmark reported leading results on document reasoning and analysis (Anthropic, CNBC).

10. A fluid simulation choreographed to music it composed in code

The most artful demo: a fluid simulation whose motion is synced to a classical-EDM remix that Fable 5 wrote in code — having, of course, never actually heard music (Anthropic). A small thing that says a lot about reasoning across domains it was never directly trained to connect.

A note on the hype: not everything viral is real

Worth saying plainly: in the first 48 hours, at least one widely-shared "look what Fable built" showcase — slick clones of Minecraft and Windows — was reported by 36Kr as possibly handcrafted rather than model-generated. Treat unsourced viral clips with skepticism. The examples above are anchored to first-party announcements, named companies, or reputable outlets for exactly that reason — and that habit, sourcing your claims, only matters more as the demos get more spectacular.

What this means if you're building something

Strip away the spectacle and one pattern runs through every credible result: the model does its best work when it is handed a complete, well-specified goal and the room to plan, build, and check itself — not when instructions arrive piecemeal. That is the same thing we tell founders before a Takeoff sprint, and it is why a sharp spec is now the highest-leverage hour of any build.

If you are sizing up your own idea, our free App Idea Validator and MVP Cost Estimator are a quick first pass, and our services page covers how we turn a weekend's worth of ambition into something safe to launch. Building something and want a second set of hands? Tell us what you're working on.

For the companion piece on what this model actually costs to run, see How Claude and Codex Actually Bill Tokens in 2026.

Reported June 11, 2026, two days after launch. Figures come from the linked sources — many are vendor demos and early reports, and claims will be tested and revised as more people build. Treat this as a snapshot, not a spec sheet.