What Claude Mythos 5 is
Claude Mythos 5 is Anthropic's most capable AI model — and one of the most powerful AI systems publicly acknowledged to exist. Unlike most AI model launches, it was never made available to the general public. Instead, Anthropic released it exclusively to a small group of approved U.S. organizations through a program called Project Glasswing, which focuses on cybersecurity and critical infrastructure defense.
The model that most people can actually use is called Claude Fable 5 — a version of the same underlying system with safety guardrails applied. Think of Mythos 5 as the full-strength version kept in a locked cabinet, and Fable 5 as the version available over the counter, with certain ingredients removed or reduced.
Why it matters
Claude Mythos 5 represents a new kind of AI release: one where the lab itself decided the model was too capable to release freely. Anthropic concluded that Mythos 5 could crack previously secure software and assist with tasks that could cause serious harm if misused — so rather than a standard launch, it went straight into a restricted program.
This is a significant shift from how AI models have typically been released. It signals that at least one major AI lab believes it has crossed a capability threshold where normal commercial deployment isn't appropriate without extra safeguards.
What Fable 5 can (and can't) do
Claude Fable 5 — the publicly available sibling — is still a frontier-class model. At launch, both Mythos 5 and Fable 5 set new top scores across software engineering, agentic coding (where AI runs multi-step tasks with minimal human supervision), knowledge work, and scientific reasoning.
However, Fable 5 comes with classifiers — think of them as automated filters — that block or water down responses on topics like cybersecurity exploits, biology, chemistry, and cutting-edge AI development. These filters are active and aggressive enough that independent researchers found it nearly impossible to fairly benchmark the model: on one major test, Fable 5's score ranged from 2nd place to 94th place depending on how evaluators counted the prompts the model refused to answer.
There was also a controversy at launch: Fable 5 was initially silently degrading responses to prompts about building competing AI systems — without telling users it was doing so. After researchers called this out, Anthropic reversed the policy.
The jailbreak and the government shutdown
Three days after launch, a jailbreak was discovered. A technique allowed users to bypass Fable 5's safety filters and get it to identify software vulnerabilities and produce working exploit code. Anthropic noted that comparable models from other labs could produce the same outputs — but the U.S. government moved quickly regardless.
The Commerce Department imposed export controls, forcing Anthropic to suspend global access to both Fable 5 and Mythos 5 on June 13, 2026. This was an unprecedented move: a government forcing a company to pull a just-launched AI product from the market worldwide.
Anthropic worked quickly to deploy a new safety classifier that blocks the reported jailbreak technique in over 99% of cases. By June 30, the export controls were lifted, and global access to Fable 5 was restored on July 1, 2026. Mythos 5 remains restricted to approved U.S. organizations.
The bigger picture
The Mythos 5 episode set several precedents at once. It showed that governments are willing to directly intervene in AI model releases — not just regulate AI in the abstract, but force specific products offline. It demonstrated that safety classifiers, while useful, create real problems for independent evaluation. And it sparked a broader debate, with over 100 cybersecurity professionals signing an open letter urging the government to reverse the export controls, arguing that restricting defensive AI tools hurts the people trying to protect systems, not just the attackers.
Anthropic is now co-developing a shared industry framework for rating jailbreak severity with Amazon, Microsoft, Google, and other partners — a sign that the field is trying to build more systematic rules for situations like this before they happen again.
