Anthropic launches its first Mythos-class model with strict safety limits

Anthropic has released Claude Fable 5, a public version of its new Mythos-class model, but with strict safeguards that block queries related to cybersecurity, biology, and chemistry. Sensitive prompts are automatically redirected to the older Opus 4.8 model, and the system aggressively filters jailbreak attempts after extensive red-team testing. Anthropic says the restrictions are necessary because Mythos 5 shows significantly stronger capabilities in areas like exploit generation and agentic hacking. Only vetted professionals in Project Glasswing will access the unrestricted Mythos 5 model, while enterprise users can use Fable 5 at premium token prices.

Read the full story on Ars Technica →