Claude Fable 5 Was Banned for 18 Days. Here's What Actually Happened — and Why Mythos Is Different.

On June 9, Anthropic released Claude Fable 5 to the public. On June 12, the US Department of Commerce applied export controls and gave Anthropic 90 minutes to comply. With no reliable way to verify the nationality of every user in real time, Anthropic made the only decision it could make in 90 minutes: it pulled both Fable 5 and Mythos 5 offline for everyone, globally, including its own non-US-citizen employees. Today, July 1, Fable 5 is back. Here's the full story of what happened, why it matters, and what the difference between Fable and Mythos actually is.

The trigger: a jailbreak nobody could agree was dangerous

Amazon researchers found a technique — a specific way of prompting Fable 5 — that got the model to identify software vulnerabilities. In one case, the model produced code demonstrating how to exploit one of them. The US government treated this as a national security event. Anthropic pushed back, noting that every other model they tested — Claude Haiku 4.5, Sonnet 4.6, Opus 4.6, Opus 4.7, Opus 4.8, GPT-5.4, GPT-5.5, and Kimi K2.7 — could produce the same output. The jailbreak didn't unlock anything uniquely powerful. It accessed something that weaker, widely available models could already do. The government moved anyway. Conversations between Amazon CEO Andy Jassy and the White House reportedly helped prompt the directive. The export control came down on June 12. Ninety minutes later, both models were dark. For 18 days, Anthropic negotiated. It trained a new safety classifier specifically targeting the reported technique, blocking it in over 99% of cases. Researchers from the US Department of Commerce's Center for AI Standards and Innovation tested the new safeguards and agreed they were "extraordinarily strong." The export controls were lifted on June 30. Fable 5 came back today.

Fable 5 vs Mythos 5: the same model, not the same thing

This is the part most coverage glosses over. Fable 5 and Mythos 5 share the same underlying architecture, the same 1M token context window, the same $10/$50 per million token pricing. They are, technically, the same model. The difference is the safety layer. Fable 5 ships with safety classifiers — automated AI systems that monitor requests and outputs in real time, flagging and blocking anything that looks like it could facilitate dangerous cybersecurity work. When Fable 5 refuses a request, the API returns `stop_reason: "refusal"` as a standard 200 response, not an error. The request gets rerouted — in the new post-ban configuration, to Claude Opus 4.8. Mythos 5 has no such classifiers. It is, by Anthropic's own description, capable of finding and exploiting software vulnerabilities more effectively than "all but the most skilled human security experts." It can turn freshly disclosed bugs into working exploits in under a day. It found a 27-year-old flaw in OpenBSD. This is why Mythos is not generally available. It exists inside Project Glasswing — a programme that gives a small number of vetted US organisations access to the model for defensive cybersecurity work. Following the export control standoff, Mythos 5 has been partially restored to approved US institutions. The broader international Glasswing expansion is still being negotiated with the government.

Why the distinction matters for developers

If you're building on the Claude API, the practical implications of this divide are real. Fable 5 is the model you have access to, and it's the right choice for virtually every production use case. It's the model I use to build FreeMalta — the Cloudflare Workers infrastructure, the News Beast editorial pipeline, the Gambity AI character, the partner recommendation logic. All of it runs on Claude. Most of it on Sonnet 4.6, which remains the model I reach for most often — fast enough to feel immediate, capable enough to handle complex multi-step reasoning, priced at a point that makes production automation genuinely economical. Fable 5 adds extended thinking, a 1M context window, and the ability to handle long-horizon agentic tasks that Sonnet can't sustain. For the work that requires it — deep document analysis, multi-hour autonomous agent workflows, complex code reasoning across large codebases — it's now back and available. The tradeoff Fable 5 makes is refusals. The new safety classifier is tighter than before the ban. You will encounter more false positives than you did in the first three days of Fable's existence. Legitimate coding and debugging requests that touch security-adjacent topics will occasionally be blocked and rerouted to Opus 4.8. This is the cost of the deal Anthropic made to get the model back online.

What this episode actually revealed

Three things became clear in those 18 days that weren't clear before. First, the US government has no proper process for this. A June 2 executive order created a voluntary framework for frontier model review before release. Fable 5 never went through it. When Washington wanted to act fast, it reached for export controls — a tool designed for physical technology exports, improvised onto an AI model. The result was a global shutdown of a product used by millions of people, triggered by a jailbreak that every comparable model could replicate. Second, the Chinese open-source developers won those 18 days. Every hour Fable 5 was offline was an hour where Chinese alternatives — Z.ai's models topped the AI ranking charts during the ban — were available to developers who would otherwise have been using Claude. The export control intended to protect US AI advantage inadvertently handed a competitive window to the exact competitors it was designed to contain. Third, the fix took a single targeted safety classifier. Not a fundamental capability rollback. Not a redesign. One filter, reviewed and approved by government researchers in under three weeks. The 18-day standoff was resolved by something that could have been shipped in the first 90 minutes. Whether the next frontier model goes through a structured pre-release government review — the framework Amazon, Microsoft, Google and Anthropic are now developing together — or faces another improvised export control directive, depends on whether anyone in Washington learned that lesson.

What "Fable 5 is back" actually means for your account

Before you assume full access is restored, read the small print — because the return comes with real constraints most announcements aren't being clear about. Through July 7: Fable 5 counts toward up to 50% of your weekly usage limits on Pro, Max, Team and select Enterprise plans. The other 50% stays on Sonnet and Opus. You don't get to choose which requests go to Fable 5 — the allocation is automatic. After July 7: Fable 5 moves to usage credits. It's no longer included in your base plan. You'll need to purchase extra credits to use it, or your requests will default back to the standard model tier. The silent rerouting problem: When Fable 5 refuses a request — including false positives on legitimate coding or debugging work — the API automatically reroutes that request to Claude Opus 4.8. You receive a notification, but if you're not watching your API responses carefully, you may not notice you've been switched. You could be building on the assumption that Fable 5 handled a request when Opus 4.8 actually did. For production integrations, this means three things you need to handle: new response logic for `stop_reason: "refusal"`, fallback configuration for the reroute, and updated billing tracking that accounts for the fact that refused requests aren't billed but rerouted ones are — on a different model's pricing. The new safety classifier is tighter than the pre-ban version. Security-adjacent coding requests, vulnerability analysis and anything touching defensive cybersecurity will hit false positives more often than before June 12. This is the price of the deal that brought Fable 5 back. For now, Fable 5 is back — with conditions. Mythos remains behind the Glasswing door. And Sonnet 4.6 is still the one doing most of the actual work.