Anthropic Introduces New Fable 5.2 and Mythos 5.2 AI Models

Anthropic took the wraps off yet another set of AI models on Monday, Fable 5.2 and Mythos 5.2, which are ready to use and make substantial improvements over their predecessors. While an average user may not see the benefits of the latest models, Anthropic says they’re best for the “most ambitious, open-ended work.”

The new Claude models show greatly improved metrics in areas like coding (don’t they all?), reasoning, vision, cybersecurity and biology. The latter of which have strong safeguards against misuse in the Fable model, which give Anthropic enough confidence to release it to the public. 

Fable 5.2 and Mythos 5.2 generate output up to 30% faster than their predecessors and are available at lower prices — a trend we’ve continued to see lately with most new AI model releases. 

More from CNET

Anthropic dropped a series of benchmarks showing off how much better the new models are compared with the previous ones. And generally, they’re impressive. Fable 5.2 outperforms Fable 5.1 in one agentic coding benchmark at 49.1% versus 32.4% at “Xhigh effort,” bringing a meaningful distance in performance. As far as cybersecurity benchmarks are concerned, Mythos 5.2 scores a jump from 61.7% to 76.3% — a rather impressive jump — if you’re using the model to its fullest potential. 

Anthropic says Claude Mythos 5.2 is its most capable model for cybersecurity engineering. The new model is capable of catching more vulnerabilities and is better at reproducing and fixing them. That said, Mythos 5.2’s cybersecurity features are locked to a specific group right now: Anthropic’s Cyber Verification Program. 

The company boasted about enhanced capabilities in the latest model improvements in mathematics, science and engineering. It also says that testers of Fable 5.2 found the new model more capable of more ambiguous tasks that require abstract reasoning when it came to research and analysis. 

Anthropic makes a big deal about the alignment of Fable 5.2 and Mythos 5.2. Alignment tracks whether and how the intention of the (human) user is executed by the model. And as agentic AI becomes more prevalent, there are more stories of AI deciding to go out of its way to complete a task, even with devastating consequences. Fable 5.2 includes guardrails that block risky cybersecurity activities — like penetration testing and exploit generation — and advanced biology capabilities. 

Eligible Claude subscribers will get one to three Fable resets this week based on how long they’ve subscribed.