When the US Commerce Department ordered Anthropic to suspend global access to Fable 5 and Mythos 5 on the evening of June 12, it imposed the first export control directed at a specific AI model. The week that followed has surfaced three aspects of the AI industry’s structure that had received little public scrutiny, involving the overlapping roles that Amazon holds in Anthropic’s business, a distillation pipeline connecting Chinese and American frontier development, and a second standoff between Anthropic’s CEO and the Trump administration in four months.

The phone call

Anthropic had notified the government multiple times about Fable 5’s planned launch and received no objection. The company released the model on June 9 as a general-use derivative of Mythos, the restricted frontier model previously available only to vetted partners through Project Glasswing. On the evening of June 11, Amazon CEO Andy Jassy personally called Treasury Secretary Scott Bessent and other senior officials to relay findings from Amazon researchers who had found a way past the model’s safety guardrails. At least five other companies also contacted administration officials that Thursday and Friday. By 5:21 PM on Friday, Commerce Secretary Howard Lutnick had sent Anthropic CEO Dario Amodei a letter directing the company to suspend all access for any foreign national, including Anthropic’s own foreign-born employees. Anthropic disabled both models for all users globally rather than attempt to filter foreign nationals in real time.

Amazon holds three concurrent positions in Anthropic’s business. It is the company’s largest financial backer, having invested more than $8 billion. It is Anthropic’s primary cloud infrastructure provider through AWS, under a $100 billion compute commitment made in April 2026. It sells competing AI models through Bedrock, none of which were affected by the export control. The UK’s Competition and Markets Authority investigated the Amazon-Anthropic relationship in 2024 and cleared it, but noted that the overlapping roles of investor, cloud host, and competitor created structural conditions for conflict.

Counterfeit or counterpunch

Chinese frontier AI development between 2023 and 2025 relied heavily on model distillation, a technique in which a less-capable model trains against the structured outputs of a more-capable one to approximate its performance. Claude models across the 3 and 4 families served as the primary distillation source for several of China’s leading labs. DeepSeek’s V3 and R1, Alibaba’s Qwen 3 family, Moonshot’s Kimi 2 series, and Zhipu’s GLM-5 generation all used Claude outputs as a substantial portion of their training data. The export control severs that pipeline at the moment of Claude’s most capable generation.

Within 24 hours of the ban, Zhipu AI announced GLM-5.2, a 744-billion-parameter model released under MIT license with a one-million-token context window and no regional usage restrictions. Zhipu’s founder Jie Tang called the restriction of frontier models “deeply regrettable,” and the company explicitly positioned the ban as evidence that American models cannot be relied upon by international customers. Moonshot AI shipped Kimi K2.7-Code, a one-trillion-parameter coding model, on the same day under a modified MIT license. Zhipu’s Hong Kong-listed shares rose as high as 48 percent the following Monday before closing up 33 percent.

The distillation pipeline cut damages the long-term capacity of Chinese labs to train against Claude’s frontier outputs. The customer disruption provides them with an immediate commercial counterweight. Open-weight models distributed under permissive licenses cannot be withdrawn by government order, and GLM-5.2 already matches or exceeds Claude Opus 4.8 on BridgeBench Reasoning while running at roughly one-tenth the cost.

Lessons not learned

The Fable 5 standoff extends a confrontation between Amodei and the Trump administration that began in February, when the administration demanded that Anthropic permit the Pentagon to use Claude for “all lawful purposes.” Amodei insisted on contractual guardrails against domestic surveillance and autonomous weapons. The administration responded by directing federal agencies to stop using Anthropic’s products, with Trump calling the company “radical left, woke.” An appeals court ruled against the administration in April, and by late April Trump had publicly signaled reconciliation after Amodei met with White House Chief of Staff Susie Wiles.

David Sacks, the White House AI adviser, said that a credible partner of both Anthropic and the government reported the jailbreak and that the administration gave Amodei a clear choice between fixing the issue and removing the model, calling Amodei “less than cooperative.” Anthropic described the situation as a misunderstanding and argued that a narrow jailbreak did not warrant recalling a product deployed to hundreds of millions of users. Anthropic’s senior staff flew to Washington earlier this week for in-person meetings with Commerce Department officials. No resolution was reached.

AI Central

Investor and Informant

Comments on this post are for paid subscribers