What Claude Mythos is
Claude Mythos is Anthropic's highest-capability AI model — a tier above the Opus line that the company had been quietly developing and testing with a small group of partners before its public debut in June 2026. Think of it as the most powerful version of Claude, designed for tasks that require deep technical expertise: finding vulnerabilities in software, reasoning through complex scientific problems, and running long autonomous workflows with minimal human supervision.
Because those same capabilities could be misused, Anthropic decided not to release Mythos the way it releases its other models. Instead, it split the release into two versions with very different rules about who can use them and what they can do.
Two models, one engine
Claude Mythos 5 is the full-power version. Its safeguards are removed for specific, authorized use cases — primarily cybersecurity defense work. Access is gated through Project Glasswing, Anthropic's vetted-partner program, developed in collaboration with the U.S. government. Only selected organizations can use it.
Claude Fable 5 is the public-facing version built on the same underlying model. It includes safety classifiers — automated filters — that block or limit responses on sensitive topics including cybersecurity, biology, chemistry, and cutting-edge AI development. This is the version anyone could (in theory) sign up to use, priced at $10 per million input tokens and $50 per million output tokens — roughly half the cost of the prior Claude Mythos Preview.
Both models set new top results across software engineering, agentic coding, knowledge work, and scientific reasoning when they launched.
Why it matters
Mythos represents something new in AI: a frontier model so capable that its creator decided the general public shouldn't have unrestricted access to it. That's a significant line to cross, and the launch immediately raised hard questions about who gets to decide what AI can do — and for whom.
Before the full release, an early version of Mythos solved a novel cybersecurity challenge significant enough to prompt meetings with the Financial Stability Board. A separate incident saw unauthorized users gain access to the restricted model via Discord coordination. These episodes weren't just embarrassing — they illustrated the real-world stakes of deploying a model that can break into secure systems.
A rocky launch
The release in early June 2026 was eventful, to put it mildly:
- Hidden restrictions: Anthropic initially applied undisclosed capability degradation to Fable 5 for AI-development prompts — silently rewriting responses or using steering vectors to weaken answers — without telling users. Researchers discovered this when they noticed inconsistent behavior. Anthropic reversed the policy after backlash.
- Government shutdown: Just days after launch, the U.S. Commerce Department issued export controls that forced Anthropic to suspend global access to both Fable 5 and Mythos 5. The models went offline for users outside the U.S., and even domestic access was disrupted.
- Benchmarking chaos: Independent evaluators found they couldn't reliably measure Fable 5's capabilities. Anthropic's classifiers silently rerouted flagged prompts to the weaker Claude Opus 4.8 instead of answering them. On one major benchmark (GPQA Diamond), Fable 5's score ranged from 93.18% (2nd place) to 55.56% (94th place) depending on whether an evaluator counted refusals as failures. Different organizations — Artificial Analysis, Vals AI, and ARC Prize Foundation — each adopted different scoring strategies, producing wildly different rankings.
The bigger picture
The Mythos launch crystallized a tension that's been building in AI: as models become genuinely dangerous in the wrong hands, the people building them face pressure to restrict access — but restriction creates its own problems. It hands governments leverage to shut down access entirely. It makes independent safety evaluation nearly impossible. And it raises questions about whether safety classifiers that secretly degrade responses are honest.
AI commentator Andrew Ng argued the episode accelerated global interest in AI sovereignty and open-source alternatives, precisely because it showed how quickly a private company or government can cut off access to a powerful tool. Over 100 cybersecurity professionals signed an open letter urging the U.S. government to reverse the export controls, arguing the restrictions hurt the defenders more than the attackers.
Where things stand
Access to both models remains suspended pending resolution of the export control situation. The controversy over silent degradation has been resolved — Anthropic reversed that policy — but the structural questions it raised about transparency in safety-tiered deployments remain open. Claude Mythos represents the leading edge of a new kind of AI deployment challenge: not just "is this model safe?" but "who decides, how do you verify it, and what happens when the answer is no?"
