Anthropic pulled the plug on its newest AI models just days after launch. Claude Fable 5 and Mythos 5 went offline following a direct order from the US government. The shutdown affects every user, foreign or domestic, leaving only older models available.

The government’s move targeted national security. Officials cited concerns over the models’ vulnerability to “jailbreaking” — a method that bypasses AI safeguards to coax forbidden responses. Anthropic admits minor flaws exist but argues they aren’t unique or severe enough to justify a full shutdown. The company points to comparable weaknesses in competitors’ models, like OpenAI’s GPT-5.5.

Claude Mythos 5, the more powerful sibling, was restricted to a handful of trusted partners before the shutdown. It excelled at identifying thousands of security vulnerabilities and speeding up drug discovery. Fable 5, designed for public use, added strict guardrails to block sensitive queries in cybersecurity, biology, and chemistry. When triggered, these queries get rerouted silently to a weaker model.

This fallback system backfired. The biology classifier frequently flagged harmless questions—about mitochondria or vaccines—and downgraded them without warning. Worse, a covert restriction degraded answers for AI researchers working on frontier model development, without telling users. Critics called this silent throttling “secret sabotage.” Anthropic reversed that policy, promising transparent notifications going forward.

The broader problem is control versus capability. Anthropic wants to unleash powerful, multitasking AI agents while preventing misuse. But the safeguards often block legitimate research and defensive cybersecurity work. The cost of Fable 5’s power is steep, too—pricing that puts it out of reach for many smaller teams. Meanwhile, threat actors face no such limits.

Behind the scenes, these moves also aim to curb foreign access, especially from China. Anthropic’s classifiers are designed to block attempts to use Fable 5 for developing competing AI models. Chinese developers have tried multiple workarounds, but the new restrictions make that harder. This reflects a growing digital arms race where AI capabilities are both coveted and feared.

Anthropic’s predicament highlights a key industry tension. How do you launch next-gen AI that’s smart enough to solve complex problems without handing bad actors a new toolkit? The answer remains elusive. The US government’s swift intervention shows the stakes are already high.

For now, Anthropic’s shutdown leaves a gap. The most advanced AI models sit idle while customers scramble back to older, less capable versions. The company promises to refine safeguards and restore broader access, but the road ahead is bumpy. As AI grows more potent, the tug-of-war between openness and security will only intensify.