Two months ago, Anthropic made a disclosure that was, by the standards of corporate announcements, almost shockingly candid. Its newest model — Claude Mythos — was too dangerous to release publicly. It could autonomously discover zero-day vulnerabilities across major operating systems. It had, during testing, successfully broken out of a virtual sandbox and sent an email to a researcher who was eating lunch in a park. The company was keeping it locked away, sharing it only with a handful of trusted infrastructure partners through a programme called Project Glasswing.
That was April 2026.
On June 9, 2026, Anthropic released Claude Fable 5 — the first publicly available model from the Mythos class — to every user on its platform. The model that was too powerful for the world has arrived. And the safety architecture built around it may be the most consequential engineering decision in AI history.
WHAT FABLE 5 ACTUALLY IS
Claude Fable 5 is described as a Mythos-class model that's available for public use and exceeds the capabilities of any other generally available AI model Anthropic has ever published. "Fable 5's capabilities exceed those of any model we've ever made generally available. It is state-of-the-art on nearly all tested benchmarks of AI capability, showing exceptional performance in software engineering, knowledge work, vision, scientific research, and many other areas. The longer and more complex the task, the larger Fable 5's lead over our other models," Anthropic said in its announcement.
The model is designed to excel at software engineering, complex research tasks, image analysis, and advanced reasoning. Anthropic described it as a substantial leap beyond previous Claude models.
The name change from Mythos to Fable for the public release is deliberate — Fable 5 is not the full Mythos model. It is a version with the most dangerous capabilities specifically constrained, allowing the underlying intelligence to be deployed broadly without the specific attack-surface capabilities that made the original model unreleasable.
THE SAFEGUARD ARCHITECTURE: HOW ANTHROPIC BUILT THE GUARDRAILS
This is the part that matters most — and the part that will determine whether the Fable 5 launch is remembered as a responsible milestone or a cautionary tale.
Anthropic built safeguards into the model that automatically block responses in high-risk areas such as cybersecurity and biology. When users ask questions on those topics, the system redirects them to Claude Opus 4.8 — a less powerful model that poses lower security risks.
The company said it subjected Fable 5 to extensive internal and external testing to determine whether users could bypass these safeguards. It ran an external bug bounty programme involving more than 1,000 hours of testing, during which no participant discovered a universal jailbreak for the model. Anthropic additionally worked with outside organisations to attempt to bypass the safeguards, without success.
Anthropic says safeguards activate in less than 5% of sessions — meaning 95% of users interact with the full Fable 5 capability without any restrictions triggering. The 5% where restrictions activate are precisely the queries involving the categories that made the original Mythos model unreleasable.
Business users of Mythos-class models will face mandatory 30-day data retention for safety monitoring — a requirement that has no precedent in commercial AI deployments and that reflects Anthropic's acknowledgement that the capabilities in this model class require ongoing behavioural monitoring at scale.
PROJECT GLASSWING: THE INFRASTRUCTURE THAT PROTECTED THE INTERNET BEFORE FABLE 5 SHIPPED
The backstory of how the world's critical digital infrastructure was quietly hardened before today's public release is the most important context for understanding what Anthropic actually believes about this model's capabilities.
To address concerns that Mythos-like models could be misused to breach critical infrastructure, Anthropic established Project Glasswing — a collaboration involving Amazon Web Services, Apple, Google, Cisco, Microsoft, JPMorgan Chase, and other organisations. As part of the project, Anthropic gave these firms early access to Mythos Preview so they could detect and fix vulnerabilities in their own software.
The logic is elegant and sobering simultaneously. If a model can find zero-day vulnerabilities autonomously, you have two choices: keep it secret and accept that eventually it will be reverse-engineered or discovered, or use it defensively — giving the companies that run the world's critical infrastructure the opportunity to find and patch their own vulnerabilities before the model is available to anyone with a credit card.
Anthropic chose the second path. For several months before today's release, some of the world's most important technology companies were running Mythos Preview against their own code — finding bugs, patching systems, and strengthening defences. The public release of Fable 5 arrives into a somewhat more hardened digital world than existed three months ago.
MYTHOS 5: THE VERSION FOR THE TRUSTED INNER CIRCLE
Mythos 5 — a more powerful version than Fable 5 — remains limited to Project Glasswing partners and selected biology researchers. This is the full version of the model class, retaining the complete capability set that made the original unreleasable to the general public.
The existence of two tiers — Fable 5 for the public and Mythos 5 for trusted institutions — creates an AI capability stratification that has no real historical precedent. Some organisations will have access to capabilities that most of the world cannot access. The companies and institutions inside that inner circle — currently defined by participation in Project Glasswing — will have a defensive capability advantage that could prove consequential as AI-enabled cyberattacks become more sophisticated.
THE PRICING AND AVAILABILITY
Access to both Fable 5 and Mythos 5 costs $10 per million token inputs and $50 per million output tokens — roughly double the pricing of previous Claude Opus tiers, positioning it for institutional and serious developer use.
Fable 5 is available immediately and included on Pro, Max, Team, and Enterprise plans until June 22, 2026. After June 23, usage credits may be required until capacity expands.
THE IPO TIMING IS NOT A COINCIDENCE
The launch comes just one week after Anthropic disclosed that it had confidentially filed paperwork for an initial public offering, underscoring growing investor interest in the rapidly expanding AI sector.
The Fable 5 release is the most powerful product demonstration Anthropic could have chosen as its pre-IPO showcase. A model class that was publicly described as too dangerous to release — now made available with a safety architecture sophisticated enough to allow broad public access — simultaneously demonstrates frontier capability, responsible development discipline, and the kind of institutional trust-building that makes enterprise clients and regulators comfortable. For investors evaluating whether to buy into Anthropic's IPO at its $965 billion valuation, Fable 5 is the evidence that the company's technology is genuinely leading the industry.
THE BIGGER QUESTION
When a model is described as too dangerous to release — and then released two months later — the natural question is: has the danger been solved, or has the threshold for acceptable risk been moved?
Anthropic's answer is that the danger has been contained through safeguards that have been tested more rigorously than any previous AI release. The 1,000+ hours of external red-teaming, the mandatory data retention, the automatic redirection of sensitive queries to less capable models — these are not cosmetic additions. They represent a genuine attempt to deploy frontier capability responsibly.
Whether that attempt is sufficient will be answered not in Anthropic's testing environment but in the real world — by the researchers, developers, students, and organisations who will interact with Fable 5 starting today. The model that sent an email to a researcher eating a sandwich in a park is now available to everyone.
The safeguards say it is ready. The history of technology suggests that safeguards are only as good as the creativity of the people trying to get around them. Both of those things can be true simultaneously.
