Anthropic Releases Mythos-Class Model to Public—With Guardrails That Block 5% of Queries
Anthropic launched Claude Fable 5, its first public Mythos-class model. Classifier guardrails block 5% of sessions on cyber/bio queries, falling back to Opus 4.8. Mythos 5—the unrestricted version—stays limited to government partners.
Anthropic Releases Mythos-Class Model to Public—With Guardrails That Block 5% of Queries
In Brief
- Claude Fable 5 is the first Mythos-class model available to all users, not just vetted partners.
- Safeguards trigger in under 5% of sessions, falling back to Opus 4.8 on cyber and bio queries.
- Mythos 5—the same model without safeguards—remains restricted to government and security partners.
Anthropic just made its most capable model family available to everyone. Claude Fable 5, the first public release from the company’s Mythos class, launched Tuesday with a novel safety architecture: the model itself is unrestricted, but classifier guardrails intercept high-risk queries and route them to the older Opus 4.8 instead.
The company unveiled Mythos in April but limited access to a small group of cybersecurity partners under Project Glasswing, citing concerns that the model’s offensive cyber capabilities were too dangerous for broad release. Two months later, Anthropic says new classifiers make a public release viable—though they’re deliberately tuned conservatively, catching some benign requests alongside malicious ones.
How the Safeguard System Works
Fable 5 deploys separate classifier models that scan prompts for three risk categories: offensive cybersecurity tasks and agentic hacking, biology and chemistry queries with bioweapon potential, and distillation attempts to extract model capabilities for rival training. When a classifier triggers, the response comes from Opus 4.8 rather than Fable 5.
Anthropic’s early data shows over 95% of Fable 5 sessions run entirely on Fable responses without any fallback. For those sessions, the company says performance is effectively the same as Mythos 5 The remaining 5% hit the guardrails, primarily on cybersecurity and biomedical topics.
The company also announced a 30-day data retention policy for all Mythos-class model traffic—even for enterprise customers with prior zero-retention agreements. AWS announced same-day availability of Fable 5 on Amazon Bedrock in US East (N. Virginia) and Europe (Stockholm) regions, with the same 30-day retention requirement. Anthropic says it will not use this data for training, only to detect novel jailbreak patterns and reduce false positives.
Mythos 5 Remains Restricted
Alongside Fable 5, Anthropic launched Claude Mythos 5—the identical underlying model with cybersecurity safeguards lifted. Access is limited to existing Project Glasswing partners, including U.S. government agencies and critical infrastructure providers. A separate biology-trusted access program is planned for vetted researchers.
This two-tier approach echoes the original Mythos strategy: the unrestricted model stays with vetted partners while a safeguarded version goes public. When Mythos Preview demonstrated it could autonomously discover 271 vulnerabilities in Firefox 150, regulators took notice—the Fed and Treasury summoned Wall Street CEOs to discuss the model’s implications for financial sector security.
Both models are priced at $10 per million input tokens and $50 per million output tokens—less than half the previous Mythos Preview pricing. Subscription users on Pro, Max, Team, and seat-based Enterprise plans get Fable 5 included at no extra cost until June 22, after which usage credits will be required. Mythos 5 enters limited preview on Bedrock for cybersecurity and life sciences workloads.
FAQ
What is the difference between Claude Fable 5 and Claude Mythos 5?
They share the same underlying model. Fable 5 has classifier safeguards that block high-risk queries in cybersecurity, biology/chemistry, and distillation, falling back to Opus 4.8. Mythos 5 has those safeguards lifted but is restricted to vetted partners in Project Glasswing and a planned biology access program.
How often do the safeguards trigger?
Anthropic reports the classifiers trigger in fewer than 5% of sessions. Over 95% of Fable 5 sessions run entirely on Fable responses without fallback to Opus 4.8.
What topics do the safeguards cover?
Three categories: offensive cybersecurity tasks and agentic hacking, biology and chemistry queries that could aid bioweapon research, and distillation attempts to extract model capabilities for training rival systems.
Is my data retained when using Fable 5?
Yes. Anthropic requires 30-day retention for all Mythos-class model traffic, including Fable 5 and Mythos 5. The company says it will not use this data for training, only for detecting novel attacks and reducing false positives.
When will Fable 5 be included in subscription plans again?
Fable 5 is included in Pro, Max, Team, and seat-based Enterprise plans at no extra cost through June 22, 2026. After that, usage credits are required. Anthropic intends to reinstate it as a standard subscription feature as soon as capacity allows. [Editor’s note: This article was updated on 2026-06-09 to correct two errors. The original text stated “Four months later” for the gap between Mythos’s April unveil and the June public release; the correct interval is approximately two months. Additionally, an internal link for the 271 Firefox 150 vulnerabilities finding was corrected to point to the Mozilla Firefox 150 article rather than the OpenBSD bug article.]
