Anthropic Opens Mythos-Class Intelligence to the Public — With a Classifier Standing Guard

Anthropic has released Claude Fable 5, a publicly available model built on the Mythos architecture the company spent months keeping on a tight leash. A second model, Claude Mythos 5, exists alongside it — identical under the hood but with the cybersecurity restrictions removed, reserved for a small cohort of vetted defenders through something Anthropic is calling Project Glasswing.

The pitch for Fable 5 is straightforward: Mythos-class reasoning for general users, minus the parts that worried Anthropic enough to restrict the original model to roughly 50 organizations back in April. The company has since nudged that number to 150. Now it's opened the gates wider, betting that classifiers can do what access controls used to.

Those classifiers work by intercepting requests that touch cybersecurity, biology, chemistry, or model-distillation topics and routing them to the less capable Claude Opus 4.8 instead. Anthropic says the fallback fires in fewer than 5% of sessions. Framed differently: most users, most of the time, are talking to a full Mythos-class model without restriction.

Here's where it gets interesting. Rob T. Lee, chief AI officer at SANS Institute, says his routine incident-response and forensic workflows triggered the redirect during initial testing. Benign, bread-and-butter defender tasks — rerouted. That's a textbook overfiring classifier problem. It's the content-moderation equivalent of blocking a search for "network exploitation" because the word "exploit" appears, rather than because the intent is malicious.

Anthropologic says the conservatism is deliberate. False positives are an acceptable price while the team tunes the system. Fair enough — but if defenders running standard IR playbooks keep hitting Opus 4.8 instead of Fable 5, the product story gets complicated fast. The whole point of Mythos-class capability for security teams is the ceiling, not the floor.

On the capability side, Anthropic's claims are notable. Stripe reportedly used Fable 5 to run a codebase-wide migration across a 50-million-line Ruby repository in a single day — a job the company estimated at two-plus months of manual engineering. The model also shows gains in financial document reasoning, chart interpretation, and reconstructing web application source code from screenshots.

For Mythos 5 specifically, Anthropic is positioning it as the strongest cybersecurity-capable model currently available, able to assist with vulnerability discovery, exploit development, and multi-stage offensive workflows. That's exactly what made earlier Mythos versions restricted. The company says it developed these capabilities in consultation with the US government and plans a gradual trusted-access expansion.

Anthropologic says extensive internal and external red-teaming found no broadly effective jailbreaks against the classifier layer. That's a meaningful claim, and one the research community will test aggressively in the coming weeks.

Cisco's Anthony Grieco offered the grounded take: access to a powerful model is not the same as operationalizing it well. Patching, MFA, network segmentation — none of that disappears because Fable 5 can migrate a Ruby monolith overnight. The AI raises the ceiling. The fundamentals determine whether anyone benefits from the extra height.