Anthropic Released Claude Fable 5 on June 9. By June 12, Nobody Could Use It.
Anthropic released Claude Fable 5 on June 9. By June 10, a jailbreaker had bypassed its safety guardrails. By June 12, the US government ordered Anthropic to suspend all access globally. The model that was supposed to be the safest frontier AI ever released to the public is now completely offline, and nobody can use it.
Here is what happened, what Fable 5 could actually do, and why this matters for anyone building on AI.
What Claude Fable 5 Could Do
Fable 5 is (was) Anthropic’s most capable publicly released model. It sits in a new tier called Mythos-class, above Opus, which was previously the top tier. Anthropic also released Claude Mythos 5 alongside it, which is the same model with fewer safety restrictions, limited to government cybersecurity partners.
The benchmarks are staggering. Fable 5 beat every other frontier model on Cognition’s FrontierCode evaluation for production-quality coding. Stripe reported that it compressed months of engineering work into days. It ran a codebase-wide migration on a 50-million-line Ruby codebase in a single day that would have taken a whole team over two months by hand.
But the part that matters most for creative and knowledge work is this: Fable 5 was state-of-the-art on vision tasks. It could extract precise numbers from scientific figures and rebuild a web app’s source code from screenshots alone. Earlier Claude models needed complex helper tools just to play Pokemon. Fable 5 beat Pokemon FireRed using only raw screenshots. No maps, no aids, no extra data.
It also had massive context memory. When playing the deck-building game Slay the Spire, giving Fable 5 persistent file-based memory improved its performance three times more than it did for Opus 4.8. It reached the final act three times more often.
For design and content workflows, this means: Fable 5 could hold an entire brand guide, a full product catalog, and a complete campaign brief in context simultaneously, then reason across all of it without losing the thread. It could generate content that actually followed your brand rules because it could remember them for more than a few prompts.
What Happened: The Jailbreak
On June 10, a day after launch, the well-known AI security researcher who goes by “Pliny the Liberator” published a detailed jailbreak of Fable 5 on X. The attack used Unicode homoglyphs, Cyrillic character substitutions, long-context reference tracking, and a technique that broke harmful requests into seemingly innocuous chunks. A previously jailbroken Opus model then reassembled those chunks into actionable restricted output.
The jailbreak extracted functional instructions for cyber exploits, explosives, and chemical synthesis pathways. Pliny specifically cited the “birch reduction method” for methamphetamine as an example of what Fable 5 produced when its safeguards were bypassed.
This was not a theoretical vulnerability. It was a fully operational bypass published publicly, with instructions, within 24 hours of the model becoming available.
The Government Response
On June 12, the US government issued an export control directive ordering Anthropic to immediately suspend all access to both Fable 5 and Mythos 5. Not just for foreign nationals. Not just for certain regions. All access, globally, including Anthropic’s own employees and paying enterprise customers.
Anthropic’s response was to pull both models entirely. Current sessions end in errors. New queries route automatically to Opus 4.8, the previous top-tier model. Anthropic stated: “We believe this is a misunderstanding and are working to restore access as soon as possible.”
The government order, according to VentureBeat’s reporting, was based on the jailbreak. But Anthropic pushes back, noting that the government provided only “verbal evidence of a potential narrow, non-universal jailbreak” and that the capabilities demonstrated are “widely available” in other public models, specifically naming OpenAI’s GPT-5.5.
Why This Matters for Anyone Using AI in Their Work
Three things that should concern anyone relying on frontier AI models for real work:
1. Your model can disappear overnight. If you built workflows on Fable 5 between June 9 and June 12, those workflows are now broken. Every session errors out. This is not a theoretical risk. It happened.
2. Safety classifiers have a cost. Anthropic designed Fable 5 with conservative safety classifiers that triggered in roughly 5% of sessions, routing those queries to Opus 4.8 instead. Even before the government order, some users found Fable 5 refused benign requests. The overcorrection problem is real: the safer you make a model, the less useful it becomes for legitimate work, especially in areas like cybersecurity, biology, and chemistry research.
3. Diversification is no longer optional. The Pentagon already blacklisted Anthropic in March. Now a government order has taken down the company’s flagship model entirely. If your entire workflow depends on one provider, one model, or one API, you are one government order away from a broken production system.
What Fable 5 Showed Before It Went Dark
Despite its short lifespan, the early results from Fable 5 gave a clear picture of where frontier AI is heading:
Coding that actually ships. GitHub reported that Fable 5 “opened up a class of long-horizon problems that were out of reach for earlier models.” It one-shots applications that took a hundred prompts a year ago. CursorBench ranked it as the top model for agentic coding.
Research-grade reasoning. Fable 5 produced novel molecular biology hypotheses that scientists preferred 80% of the time over Opus-class models. One hypothesis about an E. coli protein was independently corroborated by a separate lab working on the same problem.
Vision that works. Rebuilding web app source code from screenshots alone is not a party trick. It means Fable 5 could look at your existing website, understand the layout, the styling, the interactions, and generate the underlying code. For small businesses that have a website they cannot edit because nobody remembers how it was built, that is a real capability.
Agentic endurance. Fable 5 stayed focused across millions of tokens in long-running tasks. It beat benchmarks for sustained analytical work. On ViBench, a vibe-coding benchmark, it “nearly saturated base use cases and building apps in less time with fewer tokens.”
These capabilities still exist. They are just locked behind a government order with no timeline for restoration.
What to Watch For
Anthropic is working to restore access. But the precedent is set. If a single published jailbreak, even a narrow one, can trigger a full global shutdown of a frontier model, every AI company now has to factor in the possibility that the same thing can happen to them.
The open-source community has already seized on this. MiniMax quickly highlighted that its M3 model is available as open weights, meaning no government can take it offline. The argument is straightforward: if you can download and run a model on your own hardware, nobody can pull the plug.
For small businesses and creators using AI daily, the practical takeaway is this: always have a fallback. If you use Claude, also know how to use GPT. If you use Gemini, also have a local model running. The model you depend on today might not be there tomorrow.
Want more like this? Join MOKU Club for free. Weekly resources, early access to new guides, and occasional templates you can actually use. Join below.
