Claude Fable 5: The Model That Built Games, Migrated Stripe, Then Got Pulled

An Anthropic engineer types one sentence — "make me a self-aware version of Snake" — and a few minutes later a playable 3D game is running in the browser. Then, days later, the same model goes dark for everyone on Earth. Here is what actually happened, claim by claim.

What it felt like

For the people who got early hands on it, Claude Fable 5 felt less like a chatbot and more like a tireless junior team that never asked for a break. A Wharton researcher said he handed it a single prompt and watched it build a working game, then ran it unattended "for up to a dozen hours," grinding through multi-page instructions without losing the thread [The Shortcut; Every]. That feeling — I can walk away and it keeps going — is the whole story of this release. So is the part where it suddenly stopped.

The fair test: what Anthropic claimed

Anthropic announced Claude Fable 5 (and a restricted sibling, Claude Mythos 5) on June 9, 2026 [Anthropic]. The company's own framing was unusually direct: Fable 5's capabilities "exceed those of any model we've ever made generally available," reaching "state-of-the-art results on nearly all" tested benchmarks, with the biggest leads on long, complex tasks [Anthropic].

Games from a prompt. Anthropic and outside reviewers showed Fable 5 generating playable, browser-based 3D games from a short description — a "Library of Babel" explorer and a self-aware Snake among the demos [The Shortcut; CryptoBriefing]. It could also play games on its own: Anthropic says it finished Pokémon FireRed through a "vision-only harness" with no maps, and ran Factorio autonomously, building automated factories [Anthropic].

The Stripe migration. The headline enterprise claim: on a 50-million-line Ruby codebase, Fable 5 performed a codebase-wide migration "in a day that would otherwise have taken a whole team over two months by hand" [Anthropic]. Stripe's Matt Colyer called the results "the strongest results of any Claude model we've had the opportunity to test" [Anthropic, attributed quote]. Note the honest unit: this was a migration sped up, not a company rebuilt — "months of engineering into days" [Stripe, via Anthropic].

Long-horizon autonomy. Cognition's Scott Wu said Fable 5 was the "highest-scoring model on FrontierBench," excelling at "long-horizon reasoning"; on Cognition's FrontierCode test it "scores highest among frontier models, even at medium effort" [Anthropic, attributed]. GitHub's Mario Rodriguez said it handled "complex, long-horizon coding tasks" with autonomy that "exceeded previous benchmarks" [Anthropic, attributed]. Anthropic credits a 1-million-token context window plus note-taking that lets it "improve its outputs using its own notes" [Anthropic; every.to].

Safety routing. Here is the design choice worth understanding. Anthropic shipped Fable 5 so that, on sensitive topics, your query is quietly answered by a different model — "our next-most-capable model, Claude Opus 4.8" — rather than by Fable 5 itself [Anthropic]. The company says this fallback happens, "on average, in less than 5% of sessions" [Anthropic]. Mythos 5 is described as the same underlying model with some of those guardrails lifted, deployed narrowly through "Project Glasswing" with the U.S. government and select biology researchers [Anthropic].

The record: pricing, and the part that surprised everyone

Both models were priced at $10 per million input tokens and $50 per million output tokens — which Anthropic notes is "less than half the price of Claude Mythos Preview" [Anthropic]. From June 9 through June 22, Fable 5 was bundled into Pro, Max, Team, and seat-based Enterprise plans at no extra cost [Anthropic].

Then it was pulled. Reporting and Anthropic's own notices describe a U.S. government export-control directive, citing national-security authorities, ordering Anthropic to suspend access for any foreign national "whether inside or outside the United States" [AI Weekly; BankInfoSecurity]. To comply, Anthropic reportedly disabled access for all customers worldwide [AI Weekly]. InfoQ and CNBC covered both the release and the rapid suspension; one opinion piece in The Hill framed it as a "failed" release — that characterization is contested commentary, not Anthropic's word [InfoQ; CNBC; The Hill, opinion].

So the same week produced two true things at once: a model that compressed two months of migration work into a day, and a model the public could no longer reach. Both belong in the record.

Why it matters

Strip the hype and the durable signal is long-horizon autonomy — a model that can run for hours, keep its own notes, and finish multi-step work with less hand-holding. The Stripe number is concrete and attributable; the game demos are real but easy to over-read (a generated Snake is not a shipped studio title). And the suspension is the quiet lesson: frontier capability now collides with export policy fast enough that "available everywhere today" can become "available nowhere" within two weeks.

Kooky-sounding until proven; here, the receipts — benchmarks, a named Stripe migration, named executives, a pricing sheet, and a government directive — are unusually well-documented for an AI launch.

Note on imagery: The image is the official Anthropic wordmark logo from Wikimedia Commons (`File:Anthropic_logo.svg`), tagged public domain as a simple-shapes/text mark below the threshold of originality. We could not independently confirm the exact direct file path resolves, so it is flagged for verification.

Sources: Anthropic — "Claude Fable 5 and Claude Mythos 5"; CNBC; InfoQ; AI Weekly; BankInfoSecurity; The Shortcut; The Hill (opinion).