Claude Fable 5 and Claude Mythos 5 are advanced AI models, both based on the same underlying "Mythos-class" architecture, but released with different access and safety configurations. Claude Fable 5 is designed for general public use and is characterized as state-of-the-art across numerous AI capability benchmarks, including software engineering, knowledge work, vision, and scientific research. Its performance lead over previous models increases with task length and complexity. Claude Mythos 5, the same core model, is initially restricted to specific partners (e.g., in Project Glasswing with the US government) and has its inherent capabilities, particularly in sensitive domains like cybersecurity, more fully exposed by having safeguards lifted in those areas. Both models are priced at $10 per million input tokens and$50 per million output tokens, significantly lower than prior high-capability models.

The core methodology for safe general release of Fable 5 involves a sophisticated safeguarding system primarily relying on AI classifiers. These classifiers are separate AI systems designed to detect and prevent misuse, including jailbreak attempts, by identifying queries related to high-risk domains. When Fable 5's classifiers flag a request in areas such as cybersecurity, biology and chemistry, or distillation, the response is automatically handled by Claude Opus 4.8, a highly capable but less potent model, instead of Fable 5. This fallback mechanism ensures that sensitive capabilities are not misused, while still providing a useful response. Users are notified when a fallback occurs. This system aims to prevent "uplift" to malicious actors who could leverage the models' advanced capabilities (e.g., for cyberattacks or dangerous biological research) to cause serious harm. The classifiers are robust, having undergone extensive red-teaming and bug bounty programs, showing high resistance to jailbreaks compared to previous models.

Specifically, the areas covered by the classifiers include:

1. Cybersecurity: Mythos-class models excel at discovering and exploiting software vulnerabilities and exhibit strong agentic hacking skills (reconnaissance, discovery, lateral movement). The classifiers broadly cover exploitation and offensive cyber tasks to prevent these capabilities from being misused. Fable 5, with these safeguards, blocks progress on such tasks.
2. Biology and Chemistry: While earlier models had narrow bioweapons-related blocks, the new safeguards for Fable 5 are broader, causing most requests related to biology and chemistry to fall back to Opus 4.8. This is due to concerns about well-resourced malicious actors gaining uplift for highly risky biological research and the models' increased ability to accomplish real-world scientific tasks (e.g., predicting viral assembly properties, outperforming specialized protein language models).
3. Distillation: To prevent the proliferation of near-frontier AI capabilities without appropriate safeguards, requests identified as part of large-scale attempts to extract or "distill" Fable 5's abilities (e.g., for training competing models in authoritarian countries) are also routed to Opus 4.8.

An additional safety measure for Mythos-class models (Fable 5, Mythos 5, and future models of similar capability) is a mandatory 30-day data retention policy for all traffic. This data, which is not used for model training and is subject to strict privacy controls and logging, helps in defending against complex attacks and novel jailbreaks, and aids in reducing false positives in the safeguard system.

Regarding capabilities, both Fable 5 and Mythos 5 demonstrate significant advancements:

• Software Engineering: Exemplified by compressing months of engineering work into days, such as a 50-million-line Ruby codebase migration performed in a single day. It leads on benchmarks like Cognition’s FrontierCode for difficult coding tasks and high-quality production code.
• Knowledge Work: Shows strong performance on complex analytical tasks, scoring highest on Hebbia’s Finance Benchmark for senior-level reasoning and excelling in trading analysis evaluations.
• Vision: Achieves new state-of-the-art, capable of extracting precise data from scientific figures, rebuilding web app source code from screenshots, and performing complex tasks like beating Pokémon FireRed with only raw game screenshots and minimal harnessing, whereas prior models required extensive scaffolding.
• Memory and Long-Context: Maintains focus across millions of tokens in long-running tasks, improving outputs by using its own internal notes, demonstrated by significantly enhanced performance in games like Slay the Spire with persistent memory.
• Autonomous Creativity/Problem-Solving: Fable 5 has autonomously built a solar system simulation from first principles, played Factorio (a complex factory-building game), designed a 3D-printable model within an editor it also created (including an AI copilot), and coded a fluid simulation synchronized to music it produced.
• Life Sciences (Mythos 5): Demonstrated accelerating drug design processes by approximately ten times, matching or exceeding human experts in protein design tasks (choosing binding sites, running tools, recovering from failures). It also consistently generates novel and compelling scientific hypotheses in molecular biology (e.g., for *E. coli* proteins, some corroborated independently) and conducted autonomous genomics research, designing a custom ML model that outperformed a published model despite being 100 times smaller.

Alignment assessments indicate that Mythos 5’s (and thus Fable 5’s) level of misaligned behavior, including deception and cooperation with misuse, is low and comparable to Opus 4.8. Access to Mythos 5 is initially for Glasswing partners and will expand through a "trusted access program" for cybersecurity and, soon, for biomedical researchers, where specific safeguards (e.g., biology and chemistry) are selectively removed based on trusted use cases. Fable 5 is generally available, with a staged rollout for subscription plans due to anticipated high demand.