Observing the U.S. Government asking Anthropic to shut down its two latest models – Fable and Mythos – there is the arguably obvious take that the U.S. Government is punishing Anthropic to help OpenAI. Then there is the directive, the regulation asymmetry with GPT-5.5, the fact that Anthropic was put on Pentagon's blacklist, and the Amazon-Jassy origin detail – in other words a reactive, under-resourced enforcement that just happened to land on the company that was already in a fight with the Pentagon. No published criteria from the government, no third-party review, no written evidence, and same-day compliance forced.
On June 10, 2026, 2 days before the U.S. government ordered Anthropic to shut down access to Claude Fable 5 and Claude Mythos 5, in an essay on his personal site, Amodei asked governments for the power to block model deployments and that it needs safeguards against arbitrary or politically motivated use (read the essay here). According to Anthropic's public statement, the U.S. government directive explicitly ordered them "to suspend all access to Fable 5 and Mythos 5 by any foreign national, whether inside or outside the United States, including foreign national Anthropic employees". Because Anthropic could not differentiate foreign nationals from U.S. persons across their millions of global API users in a short window, the practical response was to shut down both models worldwide. Although asking whether this move by the Trump administration handed OpenAI a competitive edge is a good question, I believe it's the wrong first question. A better first question is to ask why the U.S. used export control law which is built for physical and software transfers to adversaries to force a capability-based product recall.
In his essay, Amodei argued that governments should have explicit authority to block deployment of an AI model if a third-party assessment finds that it poses unacceptable risk. Although the scope of threat that he set out was narrow (ex.: bioweapons, cyberweapons, loss of control) what matters more is that he attached a condition, noting that "there must be protective measures against political favoritism or arbitrary decisions". Two days later, on June 12, Secretary Howard Lutnick and the Commerce Department gave Anthropic a 90-minute compliance window based on verbal evidence of a "jailbreak" that just asks a model to patch code. Anthropic, understandably, could not verify the citizenship of every API user in 90 minutes.
U.S. export controls are designed to stop physical items or software from crossing borders to adversaries (like a specific Entity List restriction). The U.S. Government used that to force a blanket, global shutdown of the 2 models. If "finding a non-universal jailbreak that can identify software flaws" is the new threshold for a national security recall, then more than just Mythos and Fable should be pulled today. Anthropic's June 12 statement asserted that OpenAI's GPT-5.5 has comparable dual-use capability (which no independent evaluator has confirmed as of this writing). The U.S. Government set an ad hoc threshold and risks creating an unpredictable operating environment that could stifle innovation more than statutory regulation would. Since the U.S. Congress hasn't passed comprehensive AI safety legislation, the Executive branch is stretching export control law (specifically deemed exports) into a makeshift regulatory hammer.
Months of friction between Anthropic and the Pentagon preceded the ban. The Department of Defense had previously designated Anthropic a "supply chain risk" which is typically reserved for adversarial actors, after a dispute that was reportedly tied to Claude's use in military applications. Anthropic has suggested this stemmed from its refusal to allow Claude to be used for autonomous weapons and large-scale surveillance; the Pentagon disputes that characterisation. The underlying cause remains contested. This makes the favouritism narrative interesting, but not definitive because the Government could have been reactive and chosen complaint-driven enforcement. According to the Wall Street Journal, Amazon CEO Andy Jassy raised concerns with Treasury Secretary Scott Bessent that Fable 5's guardrails could be circumvented after Amazon researchers reportedly coaxed a Mythos-class model into describing restricted cyberattack methods.
While this may be in part an "OpenAI win" story, it's also an institutional design failure story where the company that called for empowering the state almost immediately got hit by the state's discretion problem it warned about, with no third-party review mechanism in place.
Whether this becomes the template or the outlier depends on what happens next; potentially litigation, the next model recall, or whether Congress ever passes legislation that makes export control law unnecessary for this purpose. The favouritism question may eventually get answered by discovery. Until then, the operative rule for any AI lab is the one Anthropic just learned: a government doesn't need a published standard to act – only a letter.