Anthropic has released Claude Fable 5.2 and Claude Mythos 5.2, its latest high-end AI models. They’re available now and are aimed less at everyday chatbot use and more at complex, open-ended work in areas such as software engineering, scientific research and security analysis.
The two models share the same underlying technology but are released with different levels of access. Fable 5.2 is the publicly available version and includes extra safeguards that restrict sensitive capabilities, particularly in cybersecurity and biology. Requests involving activities such as penetration testing, exploit generation or advanced biological work are blocked. Mythos 5.2 does not have those restrictions in the same way, which is why access to its full cybersecurity capabilities is limited to organisations approved through Anthropic’s Cyber Verification Program. Anthropic says Mythos 5.2 is its strongest model so far for security engineering, claiming it can find more vulnerabilities and do a better job of reproducing and fixing them.
The benchmark figures Anthropic has shared show significant gains. On one agentic coding benchmark, Fable 5.2 scored 49.1 percent at its highest effort setting, compared with 32.4 percent for Fable 5.1. On a cybersecurity evaluation, Mythos 5.2 increased from 61.7 percent to 76.3 percent. As with any results published by an AI company, these numbers come from Anthropic’s own testing and are measured at maximum settings, so independent evaluations will give a better sense of how much improvement developers will see in practice. Anthropic also says the models perform better in mathematics, science and engineering, and that early testers found Fable 5.2 better at handling ambiguous research tasks that require abstract reasoning.
There are practical improvements as well. Anthropic says both models generate output up to 30 percent faster than the previous versions and are cheaper to use, though it hasn’t detailed the exact pricing changes. This follows a broader trend across the AI industry, where each new frontier model is expected to be both more capable and more affordable than the last.
Anthropic is also emphasising alignment, meaning whether a model reliably follows what users actually intend. That issue becomes more important as AI agents take on more independent tasks, where a model that takes unexpected shortcuts to finish a job could cause real problems. The split between Fable and Mythos reflects Anthropic’s view that some capabilities should be limited to approved users.
For existing subscribers, Anthropic is offering one to three Fable usage resets this week, depending on how long they’ve subscribed. The larger question is whether this gated approach can scale. As security-focused AI models improve, deciding who gets unrestricted access may become one of the industry’s most difficult policy decisions.