Anthropic Launches Frontier Mythos-Class Models: Claude Fable 5 and Mythos 5

Anthropic has launched Claude Fable 5 and Mythos 5, a new class of high-performance models that excel in engineering, vision, and scientific research. Fable 5 is available for general use with conservative safeguards that fall back to Opus 4.8 for sensitive topics, while Mythos 5 is reserved for vetted partners. These models offer significant performance gains at a lower price point than previous previews, supported by new safety-focused data retention policies.

Key Points

• Anthropic introduces the Mythos-class tier with Fable 5 and Mythos 5, setting new benchmarks in coding, vision, and scientific reasoning.
• Fable 5 uses a fallback safety mechanism where sensitive or high-risk queries are handled by Claude Opus 4.8 to prevent malicious misuse.
• Mythos 5 provides unrestricted access to high-risk capabilities for vetted government and research partners through a trusted access program.
• The models show significant real-world impact, including autonomous drug design, novel genomics research, and high-efficiency software migration.
• A new 30-day data retention policy is being implemented for these models to identify and mitigate sophisticated jailbreak attempts and safety risks.

Sentiment

Overall sentiment is mixed but leans skeptical. The community is impressed by the model's apparent capabilities and many concrete user reports are enthusiastic, yet the dominant emotional tone is frustration and distrust toward Anthropic's deployment choices. HN does not reject the claim that Fable is powerful, but it strongly questions whether the product is reliable, affordable, transparent, and fair for serious technical users.

In Agreement

• First-hand users report that Fable is a major step up for difficult coding, refactoring, code review, frontend implementation, and research-style engineering work.
• Several commenters agree that the model produces cleaner, more surgical changes and needs less correction than prior Claude models on hard software tasks.
• Benchmark-focused commenters argue that the system-card and third-party results point to a real frontier advance, not merely marketing.
• Some users see conservative safeguards and trusted access as understandable for unusually capable cyber, biology, and infrastructure assistance.
• The scientific and drug-design examples are treated by some as the most important part of the announcement because they suggest practical acceleration in research.
• A few commenters accept limited retention and monitoring as a plausible safety measure for detecting novel misuse patterns and improving false positives.

Opposed

• Many users say the safeguards are too broad, routing or blocking ordinary work in security, biology, medicine, ML, hardware, and even unrelated technical conversations.
• The hidden interventions for frontier-AI development are widely criticized as anti-competitive, deceptive, and potentially a way to degrade legitimate work without telling users.
• Commenters object to paying premium rates when the system may silently fall back, steer prompts, or reduce effectiveness instead of clearly refusing.
• Pricing and temporary subscription access are viewed by many as a push toward usage billing that could make frontier assistance affordable only to wealthy organizations.
• The retention policy raises privacy, compliance, and cloud-boundary concerns for enterprise, medical, security, and regulated workloads.
• Skeptics argue that benchmark gains are hard to trust for closed-weight models and that subjective reports may be distorted by hype, marketing, or changing deployed behavior.
• Some commenters frame Anthropic's safety language as market positioning or oligopoly-building, especially because trusted partners get fuller access while ordinary users get restricted behavior.