Claude Splits in Two: One Locked Down, One for Hackers to Ignore

Here's the playbook: Claude Mythos 5 is the luxury model, strictly for organizations that've signed the paperwork in triplicate. Fewer guardrails, answers questions it probably shouldn't, and essentially means "trust me, I know what I'm doing." Claude Fable 5, on the other hand, is the version Anthropic is pushing to the masses as the "safe" one—the model that allegedly won't help you launch a cyberattack, because apparently that's something we needed to code in.

The irony is delicious: "Fable" sounds like bedtime stories for kids, yet we're talking about an AI Anthropic figures is locked down enough for regular folks. "Safe" doesn't mean "dumb," let's be clear. It means the model was trained to not hand you a step-by-step recipe for nuking a bank's network. You're still free to ask about the theory behind cyberattacks—the difference is, it won't give you the operational manual.

This split exposes a problem the AI industry keeps pretending doesn't exist: no model is truly "safe" if someone's creative enough about asking the right questions. Restrictions are guardrails, not walls. So why not just admit that Mythos is for people whose lawyers have cleared them for the unfiltered version, and Fable is for everyone else hoping nobody figures out how easy it is to work around the rules?

Antropic tries to frame it as a conscious, almost noble choice: give the trustworthy folks full freedom, give everyone else a version that at least pretends to be cautious. It's textbook corporate compromise—everyone walks away with something, or at least a ready excuse if things go sideways.

Here's the kicker: this move is basically a confession that "universal" AI restrictions don't work. If they did, you wouldn't need two different versions. Someone should say that out loud.