The Three-Week Ban That Never Ended

The ban on Anthropic's Fable 5 and Mythos 5 lasted three weeks. The compliance framework Anthropic accepted to get it lifted has no expiration date. That gap is the real product of the government's first complete AI-model enforcement cycle. On June 2, Trump signed an executive order establishing a voluntary framework for the government to assess national security risks of advanced AI models up to 30 days before public release ^[1]. The order explicitly prohibited mandatory government licensing. The framework was voluntary by law.

We’re leading China, we’re leading everybody, and I don’t want to do anything that’s going to get in the way of that lead. — Donald Trump

Ten days later, Commerce banned Anthropic's models, citing a jailbreak that could let them be used as cyber weapons ^[2]. Anthropic shut off access for all customers, not just the foreign nationals the ban targeted ^[3]. Commerce lifted the bans in stages — Mythos 5 to roughly 100 trusted organizations on June 26, general Fable 5 access on July 1 ^[3]. But the lifting was conditional. Anthropic had to accept permanent security protocols: proactive risk detection, malicious-activity reporting, and pre-release government evaluation access, covering future models, not just the ones that triggered the ban ^[3]. The mechanism is the point. The government cannot mandate compliance directly — the executive order prohibits mandatory licensing. So it uses temporary restriction as leverage: lift the ban only if the company voluntarily accepts permanent protocols. The restriction expired. The compliance framework did not. Lutnick confirmed the scope went beyond a single company:

We don't believe this kind of government access process should become the long-term default. It keeps the best tools from users, developers, enterprises, cyber defenders and global partners who need them. — OpenAI

OpenAI has already entered the same cycle. In May, it gave the government early access to GPT-5.5 for national security testing ^[4]. On June 25, GPT-5.6 was restricted to roughly 20 trusted partners, with customer-by-customer federal approval ^[5]. Sam Altman said the limited launch came at the government's request:

at the request of the US government, it is launching today in limited preview instead of the open access launch we were planning on. — Sam Altman

OpenAI resisted the permanence:

We are working with the government to get to general availability as fast as we can. — Sam Altman

Resistance may not matter. The Anthropic deal set the precedent, and the next company faces the same choice: accept permanent compliance or stay restricted. This pattern is specific to software, and the chip regime shows the contrast. CXMT, a Chinese memory-chip maker, remains on the Entity List, designated a Chinese military company. Apple's lobbying to buy CXMT DRAM faces hard political opposition and has not been granted ^[6].

Apple choosing to partner with a Chinese military company would be a grave mistake. — John Moolenaar

The chip regime uses permanent denial. The AI-model regime uses temporary denial for permanent compliance. Two factors drive the split. First, software cannot be physically interdicted at borders the way hardware can — you can stop a shipment of chips at a port, but Microsoft's Azure already sells OpenAI models to Chinese firms by routing through Singapore data centers, with ByteDance spending over $1 billion annually ^[7]. The only leverage point is the producer. Second, Chinese substitution creates pressure that makes prolonged restriction costly. Zhipu AI's GLM-5.2, trained entirely on 100,000 Huawei Ascend processors with no Nvidia chips, competes with top US models at less than one-tenth the cost ^[8]. DeepSeek V4 is priced up to 35 times cheaper than GPT-5.5 ^[9]. Chinese models now account for 17.1% of global downloads against 15.8% for US models ^[10].

On top of the base 75% off, stack an extra 90% discount for cache hits. — Victor Chen

That substitution front is context, not the stated reason for lifting — the government's stated mechanism was the security protocols Anthropic accepted. But the calculus is shaped by it: every week of ban pushes more global users toward Chinese alternatives the ban cannot reach. The compliance framework was also emerging before the ban formalized it. Anthropic's Fable 5 system card, released June 9, revealed secret restrictions that silently degraded responses to requests involving cutting-edge AI development, specifically to prevent Chinese firms from using the model for distillation ^[11]. On July 8, Anthropic will require government-issued identification and biometric face geometry for flagged users ^[12]. Neither measure has an expiration date.

The ban was three weeks. The vetting framework — proactive risk detection, malicious-activity reporting, pre-release government evaluation, identity verification — is permanent. The restriction was the instrument; the framework is the product. ^[3][12][11]

The first enforcement cycle doesn't just produce compliance from one company. It produces a template, repeatable as Lutnick put it, that makes compliance the default for every frontier AI lab. The next company — OpenAI, then Google — faces the precedent Anthropic set. The restriction was the instrument. The framework is the product.