A Government Switched Off Two Frontier Models. Your Board Will Want The Fallback Plan.

On June 13, 2026, Anthropic's customers outside the United States lost access to two of its most capable models overnight, not because of an outage or a price change but because the U.S. Commerce Department ordered it ^[4]. The company disabled Fable 5 and Mythos 5 for every non-American user, including its own foreign employees, citing an unspecified national security concern, as Fortune reported ^[5]; within weeks, OpenAI had to limit access to GPT-5.6 under similar government pressure ^[18].

That sequence changes the category of the risk. For three years the working assumption in enterprise AI has been that you rent the best model from a handful of U.S. labs and treat capability as the only variable worth optimising ^[6]. What the shutdown demonstrated, in the words of Pawel Rzeszucinski, senior director of data and AI at WebPros, writing in Forbes, is that a frontier model is not infrastructure you control but a service someone else can switch off ^[1][2]. Nothing in a commercial contract prices that. If a policy decision in Washington can revoke your inference capacity for your Frankfurt or Singapore teams, then model access sits in the same register as energy supply and connectivity, and belongs in the same part of the risk register.

The second half of the argument is what makes a fallback plausible rather than aspirational. Rzeszucinski argues the capability gap that justified dependence has nearly closed ^[7]. Zhipu's GLM 5.2 was released under an unrestricted MIT licence ^[8], and Artificial Analysis calls it "the new leading open weights model" on its Intelligence Index ^[9]. On FrontierSWE, a benchmark for long-horizon coding tasks, a review by Avenchat puts it within a single percentage point of Claude Opus 4.8 at roughly one-fifth the price ^[10], which is about 80 percent less per unit of work ^[20]. Researchers at Semgrep found GLM 5.2 with no scaffolding outperformed Claude Code on a reasoning-heavy vulnerability detection task ^[11]. The lag between open weights and closed frontier, once counted in years, is now counted in low months ^[12].

CNBC has named the resulting metric of 2026: intelligence per dollar ^[13]. Gabe Pereyra of Harvey told the network he has "been consistently surprised by how quickly the open source has caught up" ^[14]. The practical shape of this is not a wholesale migration. Cutting-edge research, hard mathematical reasoning and genuinely novel coding still favour the most expensive systems, used sparingly as a planner or reviewer rather than a default ^[19]. The volume work, document processing, classification, customer support and structured data extraction, does not need a generalist ^[15]. Research compiled by SuperAnnotate and corroborated by peer-reviewed arXiv studies finds fine-tuned small models, sometimes trained on a few hundred labelled examples, match or exceed much larger general-purpose models on narrow tasks at a fraction of the cost ^[16]. The stack stratifies: a costly frontier tier used rarely, and a cheap, owned, fine-tuned tier doing the throughput ^[17].

The operator's obligation here is unglamorous. Rzeszucinski's first recommendation is to build in-house fine-tuning capability now, on the grounds that learning it after access is disrupted is too late ^[3]. That means knowing, per workflow, which model is doing the work, what the substitute is, how long a switch takes, and who has ever tested it.