Anthropic Releases ‘Mythos-Class’ AI Model Claude Fable 5 Anthropic’s decision to release its most capable AI model yet, Claude Fable 5, just days after warning that frontier systems may soon become “too dangerous,” has intensified the debate over how fast powerful models should be deployed and how tightly they should be constrained.

From secretive preview to public release

In April, Anthropic quietly rolled out its Mythos Preview model to a small group of partners under Project Glasswing, citing concerns that its cybersecurity abilities were too risky for broad access. On June 9, the company moved from that limited experiment to a public launch, introducing Claude Fable 5 as its first “Mythos-class” system made safe for general use.

Fable 5 shares the same underlying model as the more permissive Claude Mythos 5 but is wrapped in extensive safeguards. For queries touching cybersecurity, biology, chemistry, or model distillation, it automatically hands off to the less capable Claude Opus 4.8, whose weaker reasoning is meant to blunt high‑risk assistance. Anthropic says these topic blocks are “stricter than ideal,” acknowledging that some harmless research questions will also be deflected.

Capability surge and early user tests

Across benchmarks, Fable 5 outperforms Anthropic’s previous models and rivals from OpenAI and Google DeepMind in software engineering, knowledge work, scientific research, and vision, with its lead growing on longer, more complex tasks. Independent testers are already reporting striking results: University of Pennsylvania professor Ethan Mollick found it “outperformed basically every other public model” he’d used and could work for “a dozen hours” on multi‑page specifications, generating working video games from a single prompt.

Outside observers in the AI community have echoed that enthusiasm. Researcher Andrej Karpathy called it “a super exciting release,” noting that Fable 5 is “the same underlying model as Mythos but with added safeguards” and arguing that the step change in quality is worthy of a major version bump.

Safety, access, and a two‑track future

Anthropic frames Fable 5 as a compromise between capability and control. The company says it ran extensive internal evaluations, red‑teaming, and a bug bounty program that produced “no universal jailbreaks in over 1,000 hours of testing,” before engaging external red‑team organizations that also failed to break its classifiers. Even so, it warns that an unconstrained Mythos‑class system could enable “agentic hacking” and serious real‑world harm.

To manage that risk, Anthropic is keeping the more powerful Claude Mythos 5 behind a trusted‑access program for “a small group of cyberdefenders and infrastructure providers,” while Fable 5 is rolling out broadly via API and enterprise plans, initially at a discount before shifting to higher, usage‑based pricing later in June. The split underscores a growing industry pattern: frontier models emerging first under tight control, then re‑packaged for the public with guardrails designed to keep their most dangerous capabilities just out of reach.

Continue reading https://foxvector.com/stories/019eaeab-567b-3007-71a2-2e7cb529fcb8