The Three Days of Fable 5: How the Most Powerful Public AI Model Was Released, Then Recalled

For about 72 hours in June 2026, the most capable AI model the public had ever been allowed to use was live. Then the US government ordered it taken offline. The story of Claude Fable 5 is short, but it compresses almost every live tension in the AI industry,  capability racing ahead of rules, safety built as a deliberate handicap, and a government willing to reach for export controls to pull a piece of software after release.

What is Fable 5

Anthropic launched Claude Fable 5 on June 9, 2026, calling it the first publicly available model in a new tier it labels Mythos-class, a class that sits above its existing Opus line. The pitch was not modest. Anthropic said Fable 5’s capabilities exceeded those of any model it had ever made generally available, and that it was state-of-the-art on nearly all tested benchmarks of AI capability, with particular strength in software engineering, knowledge work, vision, and scientific research. A recurring theme in the launch was duration: the longer and more complex the task, the larger Fable 5's lead over earlier models.

‍

The benchmark numbers backed the framing rather than just decorating it. On the Artificial Analysis Intelligence Index, Fable 5 launched at number one, scoring 64.9 and landing nearly five points ahead of the closest non-Anthropic model. On SWE-Bench Pro, a closely watched programming test, it solved roughly 80 percent of problems against about 69 percent for Opus 4.8 and 59 percent for GPT-5.5.

What Fable 5 could actually do

Benchmarks are abstractions; the use cases were the part that made people sit up. The single most-cited example came from Stripe, which reportedly migrated a 50-million-line Ruby codebase in a single day, work that would otherwise have taken a team around two months by hand. That kind of task compression on a job of that scale is the concrete version of the "works autonomously for longer" claim.

‍

On vision, Anthropic positioned Fable 5 as the new state-of-the-art. In the days the model was live, users put it through the kind of ambitious builds that signal real capability rather than toy demos, cloning games, and rebuilding existing software products among them. The throughline was long-horizon autonomy, staying coherent across tasks spanning millions of tokens without losing the plot.

Are Fable 5 and Mythos 5 the same model?

Technically, yes, and this is the most important thing to understand about the whole episode. Fable 5 and Mythos 5 are the same underlying model. The difference is the leash.

‍

Mythos 5 is the unrestricted version, made available only to a small group of vetted cybersecurity and life-sciences users through Anthropic's Project Glasswing. Fable 5 is that same brain with safeguards bolted on for general release, capped, or we could say, partly lobotomized. Anthropic was explicit that releasing a model this capable carries risk: without safeguards, Fable 5's abilities in areas like cybersecurity could be misused to cause serious damage. So the company built two specific kinds of restraints into the public version.

‍

First, it added interventions aimed at limiting the model's usefulness for advancing frontier LLM development itself, narrowing how much the model can be turned toward building the next generation of powerful models.

‍

Second, and more visibly to users, it built a fallback mechanism. When Fable 5 receives a prompt touching sensitive domains,  cybersecurity, biology, chemistry, health, and bloodwork-adjacent topics fall in here, it quietly hands the request to the weaker Opus 4.8 instead of answering with its full Mythos-class capability. Independent evaluators measured how often this happens: on Humanity's Last Exam, Fable 5 triggered the safety fallback on roughly 9 percent of tasks, and Artificial Analysis observed refusal-or-fallback behavior in about 9 percent of cases overall, roughly one in eleven sessions. The trade-off is disclosed by Anthropic as a design choice, but at the moment, the substitution is silent; the user isn't told that this particular answer came from a lesser model. For most people, that's invisible; for anyone building a workflow that depends on consistent capability, it's a real caveat worth knowing.

The Fable 5 shortcomings

Two stood out even before any government got involved.

‍

The first is cost. Fable 5 is token-hungry and priced to match, $10 per million input tokens and $50 per million output tokens, roughly double Opus 4.8 and about three times Sonnet 4.6, making it the most expensive generally available model Anthropic ships. The deeper problem isn't the per-token rate; it's how fast real usage compounds. Because the model reasons at length and generates long, thorough outputs, agentic and multi-step workflows can burn through hundreds of thousands of tokens before a user notices. Subscribers reported their plan limits draining noticeably faster than with Opus; some developers running it continuously for production estimated costs in the hundreds or thousands of dollars a day. The honest counterpoint, which several cost analyses make, is that the right unit is cost per finished task, and on hard work that Fable 5 actually completes, it can come out cheaper than a weaker model that needs human cleanup.

‍

The second is the fallback behavior itself, genuinely useful as a safety measure, genuinely a limitation if you needed the full model on a sensitive-but-legitimate question and got silently downgraded.

Why the government decided Fable 5 was dangerous

Here, the story turns on a single technical dispute. The chain reportedly began with Amazon, Anthropic's largest investor and primary cloud partner, and also a direct competitor for the same enterprise customers. Amazon CEO Andy Jassy told Treasury Secretary Scott Bessent and other senior administration officials that Amazon researchers had used a series of prompts to get Fable 5 to produce information useful for conducting cyberattacks. It was not clear whether Amazon ran those tests on its own or at the government's request; one account suggested the government had asked Amazon for feedback on the new model.

‍

Those conversations preceded a letter from Commerce Secretary Howard Lutnick, written with officials from the Commerce Department's Bureau of Industry and Security, invoking national security authorities. Anthropic's own read of the underlying report was far narrower: its experts concluded it referred to a limited ability to use the model to review specific program code and fix errors, a partial bypass of a narrow safeguard, not a wholesale jailbreak of the model. The White House saw it differently. AI adviser David Sacks said a trusted partner had come forward with a jailbreak of the guardrails, that the administration asked Anthropic to fix or pull the model, and that Anthropic refused, a characterization the company disputes.

What actually happened

The directive arrived on Friday, June 12, at 5:21 p.m. ET. It was an export control order, not a consumer-safety recall, and that distinction shaped everything that followed. The order required Anthropic to suspend access to both Fable 5 and Mythos 5 for all foreign nationals, a category that swept in not just overseas users but non-citizen Anthropic employees inside the United States. One account reported the company was given 90 minutes to act, with no prior warning of a national security concern.

‍

Because Anthropic could not verify the citizenship of every user in real time, the only way to comply was to shut both models down entirely, worldwide. Amazon, Google Cloud, and Microsoft Foundry revoked access in step. Anthropic's other models, including Opus 4.8, were unaffected.

‍

Anthropic complied while publicly objecting. It argued that the finding of a narrow potential jailbreak should not be grounds for recalling a commercial model deployed to hundreds of millions of people, that perfect jailbreak resistance is not currently possible, and that its safety approach was built as defense-in-depth, monitoring layered on top of targeted resistance rather than a single perfect wall. The company warned that if this standard were applied across the industry, it would effectively halt all new frontier model deployments, and said it believed the episode was a misunderstanding it was working to resolve.

Why this matters beyond one model

Three things make this larger than a single product being pulled.

‍

It is, by available accounts, the first time the US government has used export controls to halt public access to a commercially deployed AI model, the same legal machinery used to restrict advanced chips, now pointed at software after release.

‍

The conflict of interest is hard to look past. The company whose CEO triggered the ban has tens of billions of dollars invested in Anthropic, hosts its infrastructure, and competes for the same customers. Anthropic has not publicly accused Amazon of bad faith, but the structure of the situation speaks for itself.

‍

And it accelerated the debate over sovereign AI. When a model that hospitals, companies, and researchers were already depending on can be switched off worldwide by one government's order, the argument that critical AI infrastructure shouldn't be something a foreign state can unplug overnight stops being abstract.

‍

As of mid-June 2026, Anthropic had sent senior technical staff to Washington to work toward restoring access, with no announced deal and no confirmed timeline for the models' return.