# Anthropic Pulled Fable 5 and Mythos 5: What the Ban Means > Source: > Published: 2026-06-15 13:18:54+00:00 Anthropic disabled Fable 5 and Mythos 5 on June 12 after a US government export-control order The order bars foreign nationals, so every customer lost access while compliance is sorted out The trigger was a narrow jailbreak that asked the model to fix flaws in a codebase Opus 4.8 and every other Claude model still work, so that is where I moved my pipelines Six days after I started routing real work to Claude Fable 5, it vanished. On June 12 Anthropic disabled both Fable 5 and Mythos 5 worldwide, and the reason is not a billing glitch or a quiet deprecation. A US government directive forced the shutdown. Here is what actually happened, why it happened, and the calm fallback I used so none of my projects stalled. Anthropic launched Fable 5 and Mythos 5 on June 9, the first public models in the new Mythos class. Fable 5 was free on Pro, Max, and Enterprise plans through June 22, it was rolling out across platforms including AWS Bedrock, and the paid pricing was set to kick in on June 23. So a lot of people, me included, spent that week pushing it hard on coding and long-context tasks while it cost nothing. Three days later it was gone. In its public statement, Anthropic says it received a government directive on June 12 at 5:21pm Eastern. The company describes it as an export-control directive citing national security authorities. The order bars all access to Fable 5 and Mythos 5 by any foreign national, whether inside or outside the United States, and that explicitly includes Anthropic's own foreign-national employees. The practical problem is that you cannot cleanly separate every foreign national from a live consumer product on a few hours of notice. Citizenship is not a field most chat sessions or API keys carry. So the net effect, in Anthropic's own words, is that it had to abruptly disable Fable 5 and Mythos 5 for all customers to stay compliant. Not foreign users only. Everyone. If you had a chat open or an API job queued against either model, it stopped mid-flight, and the model picker simply dropped the options. One detail matters for anyone in a panic: every other Claude model is fine. The directive names Fable 5 and Mythos 5 specifically, so Opus 4.8, Sonnet, and Haiku kept running without interruption. I cover the model that absorbed my workload in [Claude Opus 4.8 Is Here](https://dev.to/blogs/lab/claude-opus-4-8-is-here-everything-that-changed), and that piece turned into my recovery plan within the hour. The trigger was a jailbreak. Governments worry that a frontier model can be talked into helping with things it should refuse, and the concern here centers on cyber capability: could someone coax Fable 5 into finding and exploiting software vulnerabilities at a level that changes the threat picture for everyone defending real systems. Anthropic's account of the demonstrated technique is narrower than the headline suggests. As the company describes it, the jailbreak involved asking the model to read a codebase and fix software flaws, and the flaws it surfaced were minor and already known. Anthropic calls it a narrow, non-universal jailbreak, language that does a lot of work. Narrow means it did not generalize into a master key for arbitrary harmful requests. Non-universal means it did not reliably reproduce across prompts and contexts. In plain terms, the company is arguing this was a corner case, not a kicked-in door. That framing lines up with how the Mythos class was sold at launch. The whole pitch was that Mythos-grade capability ships with extra safeguards, including a classifier that can fall back to a safer response on sensitive cyber and bio prompts. I walked through that safety architecture in [Claude Fable 5 Is Here](https://dev.to/blogs/lab/claude-fable-5-is-here-the-first-public-mythos-class-model). A model marketed on its safeguards getting pulled over a safeguard gap is the kind of irony that makes a story travel, and it traveled fast. I am not going to pretend I can grade the actual risk. I cannot see the red-team report and neither can you. What I can say is that a three-day-old model sitting at the absolute frontier of coding ability is exactly the kind of thing a security agency reacts to fast, and reacting fast tends to mean a blunt instrument rather than a scalpel. A blanket suspension is the bluntest instrument there is. If you built anything on Fable 5 in that first week, treat it as offline and do not wait for it to blink back on. The honest move is to assume an outage measured in weeks, not hours, and route around it today rather than refreshing a status page. For most tasks the swap is painless because the gap between tiers is smaller than the launch benchmarks implied in daily use. Fable 5 was the sharper instrument on the hardest coding problems, and it topped the toughest coding benchmarks at launch, but Opus 4.8 already handled the bulk of my real work and it never went anywhere. I moved my agent loops, my blog generation, and my code review passes straight back to it by swapping a single model name in my config. Nothing else changed. If you are weighing the two tiers for the long run, I put them head to head in [Claude Fable 5 vs Opus 4.8](https://dev.to/blogs/lab/claude-fable-5-vs-opus-4-8-is-double-the-price-worth-it), and the short version is that the cheaper tier wins more often than the spec sheet suggests, since Fable 5 listed at roughly twice the price of Opus 4.8 per token once the free window closed. There is a quieter lesson here about dependency. I had been tempted to hard-wire a few automations to Fable 5 because it was free through June 22 and genuinely strong on agentic coding. Good thing I did not. When a model can disappear on a government's timetable rather than a vendor's roadmap, you want every pipeline to fall back to a second model with one config change instead of a rewrite. Mine did, so my Monday looked completely normal while the news cycle did not. That is the entire payoff of building model choice into your tooling rather than betting a workflow on one specific name that happened to be hot that week. As of today the models are still down. Anthropic has not restored access, and it has framed the whole episode as a likely misunderstanding. The company says it is working to bring access back as soon as possible and has reportedly sent senior technical staff to Washington for meetings with White House officials to argue that the cited jailbreak does not justify a blanket ban. No firm restoration date has been given, and the reporting I trust suggests it could take weeks rather than days, with enterprise customers being told to plan around the absence rather than wait it out. My plan does not depend on any of that resolving. I am keeping production work on Opus 4.8, which is unaffected and already proven in my stack, and I am treating Fable 5 as a nice-to-have that has to re-earn trust before it touches anything that matters. I am not rebuilding around it until it is back and stable for at least a couple of weeks, because a model that came and went once can do it again on the same kind of notice. And I am watching the policy side more closely than the model side, because the precedent is the real story. A capable model was switched off by directive, not by deprecation, and that is a new variable every builder now has to price into how they choose a default model. If you only take one action from this, make it boring and useful: confirm that every automation you run can name a fallback model and switch to it without a human in the loop. I keep that wiring documented as part of my [Claude Blueprint](https://dev.to/pages/claude-blueprint), and this week is the clearest argument I have ever had for why it belongs there. Fable 5 was the most capable coding model I had ever pointed at my own repos, and it lasted six days before a government order took it offline for everyone. The cause was a narrow jailbreak involving code-flaw fixing, the scope was a foreign-national export restriction that Anthropic could only honor by disabling the models for all customers, and the current status is still dark with no promised return date. None of that touched Opus 4.8, which is why my work never stopped for a minute. The takeaway is not which model is best this month. It is that any single model can vanish on someone else's schedule, so the resilient setup is the one that shrugs and switches. If you want the version of that setup I actually run, start with the [Claude Blueprint](https://dev.to/pages/claude-blueprint) and wire a fallback before you need one.