GLM-5.2 Open Source as US Blocks Fable 5

On June 13, 2026, the fault line in the AI race moved overnight — from who owns the smartest model to who is allowed to run it.

Less than a day earlier, the US government forced Anthropic to block access to Claude Fable 5 and Mythos 5 for all foreign nationals. Hours later, China's Z.ai (formerly Zhipu) unveiled GLM-5.2 and promised a GLM-5.2 open source release next week under the MIT license. Its tagline — 'the future of AI is open, and it belongs to the people' — read as a shot at the exact moment Washington walled off its own frontier.

Two Days, Two Locks on Fable 5

Anthropic had opened that frontier just four days earlier. On June 9 it released Fable 5, the safety-hardened public version of Mythos, held back since an April preview over fears it could find and weaponize cybersecurity flaws at scale. The unrestricted twin, Mythos 5, stays limited to vetted partners through Project Glasswing.

The first lock was Anthropic's own. Fable 5 runs classifiers that flag cybersecurity, biology, chemistry — and, crucially, frontier model development, including 'distillation.' Flagged queries aren't refused; they are quietly downgraded to the older Opus 4.8. On the safeguard backlash, Brookings fellow Kyle Chan put it bluntly to the South China Morning Post.

Chinese AI developers might find it nearly impossible now to use Anthropic's latest model to accelerate their own model development.

The second lock was the government's. At 5:21 pm ET on Friday, June 12, Anthropic received an export controls directive citing national security. The export controls covered 'any foreign national, inside or outside the United States, including foreign national Anthropic employees' — so broad that Anthropic had to block Fable 5 and Mythos 5 for everyone. It called the sole concern a 'narrow, non-universal jailbreak' — read a codebase, fix its flaws — that surfaced minor, known bugs other public models find anyway. WIRED reported the order came with only verbal evidence.

GLM-5.2 Steps into the Regulatory Gap

While Anthropic withdrew its flagship, Z.ai shipped one. The Fable 5 block opened a gap, and GLM-5.2 walked through it — out the same day across every GLM Coding Plan tier, with a 1M-token context window and long-horizon agentic coding.

The counterpunch is the next line: the API, chatbot, and open weights all land next week, with GLM-5.2 going open source under the permissive MIT license. Z.ai has done this before — GLM-5 (744B parameters, trained on Huawei Ascend chips, zero NVIDIA silicon) and GLM-5.1 both went open source under MIT, and GLM-5 hit 77.8% on SWE-bench Verified, the top open-weight score then.

There are no GLM-5.2 benchmarks yet, a real caveat. But the shape is clear: as export controls block foreign access to Fable 5, the Chinese open source camp hands a frontier-grade coding model to anyone, free to self-host. The 'AI belongs to the people' slogan does the rest.

Why Washington's AI Blockade Could Backfire

Step back, and the optics for American AI aren't flattering. Washington branded its top lab a 'supply-chain risk' in February — now in litigation — then ordered that lab's best model switched off for every non-citizen, its own foreign staff included. Even CEO Dario Amodei, who backs a 'fair, transparent' process for blocking unsafe models, said this order 'does not adhere to those principles.' The signal to global developers is plain: American frontier AI is hostage to politics.

Meanwhile the open-weight challengers keep shipping — GLM, DeepSeek, Moonshot's Kimi, Alibaba's Qwen — and US export controls have arguably hardened them, not slowed them. Zhipu has sat on the US Entity List since January 2025, cut off from NVIDIA H100s, yet trained a frontier model on Huawei chips and became the first publicly traded foundation-model company.

That is the risk baked into export controls: every model the US blocks pushes developers toward a free, self-hostable Chinese alternative, and toward the 'AI belongs to the people' framing with it. Export controls can deny compute, not mindshare. Keep closing, and the closed frontier cedes the one thing it can't sanction back — the default.

GLM-5.2 still shipped without a benchmark, and 'open next week' is a promise, not a download. But the question the day leaves is no longer whose model is best — it's whose model you're allowed to run.