Anthropic Releases Claude Fable 5: The First Mythos-Class Universal Model with Safety Guardrails

Anthropic officially releases the most powerful Claude model to date: Claude Fable 5 is the first Mythos-level model (with safety guardrails) open to general users. It reaches SOTA on almost all test benchmarks in the fields of software engineering, knowledge work, vision, scientific research, etc. The more complex and longer the task, the more obvious the lead. Claude Mythos 5 is a version with no security guardrails and is only available to cyber defense organizations through Project Glasswing. An upgraded version of the security classifier was simultaneously launched, which according to official tests eliminated 95% of hazard vectors. Pricing $10/$50 per million tokens.

Online evaluation (Chatbot Arena): Fable 5 ranks first in all categories, ahead of the second place 70 ELO - equivalent to the gap from GPT-5.4 to GPT-5.5. The winning rate is as high as 86%, and the tie and match winning rate is 96%. Programming ability ELO reaches 1794, coding style task score is 96.8.

Pricing: Input $10/million tokens, output $50/million tokens.

Safety guardrail design

The capabilities of the Mythos-level model come with risks—Fable 5’s capabilities in areas such as cybersecurity could be exploited for malicious purposes. Therefore, Anthropic deployed guard classifiers (Guard Classifiers) for Fable 5, and queries for certain categories will be handled by Claude Opus 4.8. These safety guardrails are tuned to be conservative - triggering on average in less than 5% of sessions.

Upgraded security classifier

While maintaining a pass rate of more than 92% of benign prompts, the new classifier can eliminate 95% of harmful vectors in the official test data, which is a significant improvement from the previous level of 60%. The AI ​​Security Institute (AISI) independently assessed that it is sufficiently robust.

Mythos 5 + Project Glasswing

Mythos 5, which is the same underlay as Fable 5 but removes the security guardrails, is only available to a handful of cyber defense organizations and infrastructure providers through the Project Glasswing partnership program. Anthropic will be deployed in partnership with the U.S. government.

Practical application cases

• Stripe: Completed two months of traditional work in one day (for the first time, an internal security alarm was triggered due to "excessive ability")
• Evozyne Drug Design: Design new proteins 10 times faster

The Claude Fable 5 is Anthropic's most radical model release yet. Mythos-level capabilities provide universal access to the outside world through a "security classifier" downgrade mechanism, while limiting the unrestricted version to network defense scenarios - this "capability classification" strategy represents a new paradigm for cutting-edge AI deployment. The actual cases of Stripe completing two months' worth of work in one day and Evozyne's drug design speeding up 10 times illustrate that AI's capabilities in enterprise-level programming and drug development are rapidly approaching the "replacement level".

Currently DeepSeek V4 Pro, GPT-5.5 and Claude Fable 5 have formed a tripartite competition in terms of programming capabilities, but Fable 5's overall lead in the overall benchmark shows that Anthropic has established a significant advantage at the Mythos level.

Directions worthy of follow-up:

1. Actual performance of Guard Classifiers: Acceptability of 5% trigger rate in real production scenarios
2. Extensions to Project Glasswing: Will access to Mythos 5 be expanded?
3. EXPORT CONTROL IMPACT: The actual impact of U.S. government export controls on global access to the Claude family of models