Anthropic Launches ‘Mythos-Class’ AI Models Fable 5 and Mythos 5 Anthropic’s newest AI models promise unprecedented power while reviving fresh worries about how far, and how fast, frontier systems should spread.

From private Mythos preview to public Fable 5

In April, Anthropic quietly tested a highly capable “Mythos Preview” model with a small group of critical-infrastructure partners, citing concerns that its cybersecurity skills were too dangerous for broad release. By early June, the company judged new mitigations sufficient to debut Claude Fable 5, describing it as a Mythos‑class system “made safe for general use.”

On June 9, outlets including Axios, The Verge, and Business Insider reported the launch of Fable 5 as Anthropic’s first Mythos‑level model for the public, with strict safeguards that automatically reroute high‑risk cybersecurity, biology, chemistry and model‑distillation questions to the weaker Claude Opus 4.8 model. Anthropic stressed that without such controls, the underlying model’s offensive cyber capabilities “could be misused to cause serious damage.”

Simultaneously, the firm rolled out Claude Mythos 5 to vetted cyberdefenders via Project Glasswing, calling it “the strongest cybersecurity” model in the world and lifting some safeguards for that narrow group.

Capabilities and commercial rollout

Early benchmarks and hands‑on testing suggested a step‑change in performance. Anthropic and reviewers said Fable 5 is state‑of‑the‑art across nearly all tested tasks, with a widening lead on long, complex workflows such as software engineering and scientific research. AI scholar Ethan Mollick reported that the model could generate entire video games and detailed data visualizations from single prompts, working for “a dozen hours” on multi‑page specifications.

Fable 5 was initially made broadly available to Anthropic’s Pro, Max, team and enterprise users before shifting to a usage‑credits model later in June, reflecting both high demand and compute constraints.

Praise, backlash and a safety stress test

AI practitioners welcomed the leap in capability. Andrej Karpathy called it “a super exciting release,” emphasizing that Fable 5 uses the same underlying model as Mythos with “added safeguards” and appears “SOTA on everything by a margin.” Perplexity AI CEO Aravind Srinivas integrated Fable 5 as an “orchestrator” for long‑running agentic workflows, highlighting its usefulness for complex automation.

Security‑minded observers focused on the model’s guardrails. Anthropic said classifiers watch for risky topics and jailbreak attempts, with over 1,000 hours of bug‑bounty and external red‑team testing reportedly turning up no “universal jailbreaks.” The firm admitted the filters are “stricter than ideal,” meaning some benign queries are blocked, but claimed false positives in under 5% of sessions.

At the same time, competitive tensions surfaced. A widely shared post alleged Anthropic had “covertly degrad[ed] performance for competing AI researchers” using Fable 5, prompting a swift policy reversal, according to a Wired report amplified on X. Other critics framed the heavy restrictions and cloud dependence as a reason to favor local, user‑controlled models, with one viral comment declaring, “Fable is banned. Long live local AI.”

Anthropic’s own messaging underscores the contradiction: just days before Fable 5’s release, the company urged a global “brake pedal” on frontier AI, warning that systems like Mythos might soon achieve recursive self‑improvement. For now, Fable 5 and Mythos 5 are being presented as a compromise—pushing technical boundaries while betting that layered safeguards, trusted‑access programs and market demand can keep the most dangerous uses in check.

Continue reading https://foxvector.com/stories/019ec5d8-3954-07fd-7070-19e1b5ff75c8