US bans Anthropic Fable 5 globally over jailbreak fears; China's Z.ai launches GLM-5.2 open-source rival same day

On June 12, 2026, the US Commerce Department invoked national security export controls to order Anthropic to disable Fable 5 and Mythos 5 for all foreign nationals, including Anthropic's own non-citizen employees, effective within hours. The government cited a jailbreak—a bypass of Fable 5's safeguards that could expose the underlying Mythos 5 model's advanced cyber capabilities—but provided only verbal evidence and no written disclosure. Anthropic disagreed, calling the jailbreak narrow, replicable in other models like GPT-5.5, and insisting it had built defense-in-depth safeguards, but had no choice but to pull both models globally to ensure compliance.

On June 13, one day later, China's Z.ai (formerly Zhipu AI) announced GLM-5.2, a 744-billion-parameter model with a 1-million-token context window released under an MIT open-source license—unrestricted, downloadable, and freely redistributable. The timing and framing were deliberate: Jie Tang, Z.ai founder, posted that frontier AI should not belong to a few or be withdrawn "for non-technical reasons." GLM-5.2 already tops some benchmarks on reasoning (BridgeBench Reasoning 42.8) and on agentic coding tasks, with pricing roughly one-tenth of Anthropic's Claude tiers.

For architects and security teams, the sequence reshapes the entire frontier-model calculus. A closed US model restricted by government fiat at the moment of deployment; an open Chinese model released with zero regional locks within 24 hours. GLM-5.2 is not universally better than Fable 5 on every dimension, but it is available everywhere, costs less, and is licensed under permissive open terms—meaning no single vendor or government can revoke it. The episode signals that export control is no longer a reliable tool for protecting frontier capability gaps when substitutes can be released at internet speed.