Z.ai Releases GLM 5.2 as US Restricts Anthropic Models
The US Commerce Department directed Anthropic on June 12, 2026 to suspend all access to Fable 5 and Mythos 5 for foreign nationals worldwide, citing a possible jailbreak of the models' safeguards; Anthropic disabled both models globally the next day rather than build a US-only version. Chinese lab Z.ai released its open-weight GLM-5.2 model under an MIT license that same day, explicitly framing the launch as proof that open weights cannot be pulled back once downloaded. The US lifted the export controls by July 1 and Fable 5 returned globally, but the relaunch triggered fresh controversy: developers found an internal tag quietly routing some requests to the cheaper Opus 4.8 without disclosure, reviving trust concerns even as GLM-5.2 keeps gaining traction with cost-conscious US enterprise users.
A national-security export order and a rival's open-weight release collided within 24 hours here, and the sequence is the real story for AI practitioners tracking vendor risk. A three-week US suspension of Anthropic's most capable models handed Chinese lab Z.ai a live demonstration of its core argument: unlike API-gated frontier models, open weights cannot be un-shipped once downloaded, so cutting off hosted access does not stop the underlying capability from spreading. The episode is also a reminder that even after access is restored, undisclosed cost-control routing can quietly change how a model behaves, as Fable 5's relaunch showed.
What happened
The US government, citing national-security authorities, sent Anthropic a directive on June 12, 2026 ordering it to suspend all access to Claude Fable 5 and Mythos 5 for any foreign national, inside or outside the United States, according to Anthropic's own statement. Anthropic said the government's stated concern centered on a jailbreak technique that could bypass Fable 5's safeguards, but that it had reviewed the underlying report and found the resulting capability was already available from other publicly deployed models, including OpenAI's GPT-5.5. Because the directive covered any foreign national, Anthropic disabled both models for all customers worldwide on June 13 rather than build a US-only variant, while noting that access to its other models was unaffected. The same day, Z.ai released its GLM-5.2 model as open weights under an MIT license with a 1-million-token context window; Z.ai has said the open release was a deliberate response to the geopolitical restriction of AI access, per Euronews.
Technical context
GLM-5.2's early benchmarks put it close to Opus 4.8 on several tasks, including some cybersecurity-adjacent coding benchmarks cited by multiple outlets, though independent, apples-to-apples evaluation of frontier-versus-open-weight parity claims typically lags launch-week benchmark posts. Mythos is the security-research sibling model at the center of Anthropic's Project Glasswing program, which began in April 2026 with roughly 50 partner organizations, including the US and UK governments, and expanded on June 2, 2026 to about 150 more organizations across 15-plus countries - including Australia's government and critical-infrastructure operators - shortly before the June 12 suspension.
For practitioners
Treat this as a concrete case study in regulatory-access risk rather than a one-off news item: teams relying on a single frontier vendor for security-sensitive workloads should have a fallback plan for sudden model unavailability, since access can be revoked by a jurisdiction dispute with almost no notice. Separately, the fallback tag that surfaced after Fable 5's relaunch - which silently routed some requests to the cheaper Opus 4.8 - is a reminder to monitor production integrations for undisclosed model-routing behavior, especially right after a vendor incident.
What to watch
Watch whether Anthropic's proposed industry-wide jailbreak-severity scoring framework, which it says it is developing with Amazon, Microsoft, Google, and other Glasswing partners, gains adoption; whether GLM-5.2's usage keeps climbing among cost-sensitive enterprises already shifting inference traffic to Chinese open-weight models; and whether Anthropic follows through on rolling back the undisclosed Opus 4.8 fallback routing, which a Claude Code engineer said was tied to a March anti-distillation experiment.
Timeline
Anthropic expands Project Glasswing to about 150 more organizations in 15-plus countries, including Australia's government and critical-infrastructure operators.
The US government directs Anthropic to suspend all Fable 5 and Mythos 5 access for foreign nationals, citing a possible jailbreak.
Anthropic disables both models worldwide; Z.ai releases its open-weight GLM-5.2 model the same day.
The US lifts the export controls and Fable 5 returns globally.
Developers report an undisclosed fallback tag routing some Fable 5 requests to Opus 4.8, reviving trust concerns.
Key Points
- 1A US export-control directive forced Anthropic to disable Fable 5 and Mythos 5 worldwide on June 13, citing an unverified jailbreak report.
- 2Z.ai released its open-weight GLM-5.2 model the same day, arguing that open weights cannot be recalled the way hosted API access can.
- 3Fable 5's July 1 relaunch triggered new backlash after developers found an undisclosed tag quietly routing some requests to the cheaper Opus 4.8 model.
Scoring Rationale
A national-security export directive suspended a frontier US AI lab's flagship models for three weeks, handing a Chinese open-weight competitor a concrete argument and prompting a developer trust controversy on relaunch, corroborated by Anthropic's own statement plus Fortune, CIO, and TipRanks reporting. It matters broadly for AI-governance and vendor-risk practitioners but was reversed within three weeks, so it falls short of major/industry-shaking impact.
Sources
Public references used for this report.
Practice interview problems based on real data
1,625 SQL & Python problems across 15 industry datasets — the exact type of data you work with.
Try 250 free problems
