The Trump administration’s restrictions on the leading US AI models are increasingly being described by critics as an own goal that benefits China. On Friday, after a two-week shutdown ordered under an export-control directive, the White House let Anthropic restore its powerful Mythos 5 model for roughly 100 vetted companies and federal agencies, though its consumer-facing Fable 5 model remains blocked while talks continue.
The same day, OpenAI said it would limit the rollout of its new GPT-5.6 models to about 20 government-approved partners, also at Washington’s request. The government argues the controls are needed because such models could accelerate sophisticated cyberattacks.
The timing is what has alarmed parts of the industry. Just as US labs throttle access, Chinese developers are shipping capable models with no such limits. Researchers say Zhipu AI’s GLM-5.2, an open-weight model released in mid-June under a permissive license, performs on par with top US labs on some cybersecurity benchmarks.
In tests by the security firm Semgrep, GLM-5.2 outscored Anthropic’s Claude Opus 4.8 on several vulnerability checks, and researchers said added prompting lifted it to Mythos’s level at spotting exploitable bugs, even as it trails US leaders on broader reasoning. A Jefferies strategist told clients GLM-5.2 is nearly a match for Anthropic in the corporate market at roughly a quarter of the cost per token.
Prominent voices have framed the contrast as a policy failure. Venture capitalist Marc Andreessen called GLM-5.2 the first Chinese model to match or beat US labs “with no compromises,” noting the timing. Georgetown researcher Sam Bresnick called the moment a wake-up call. Even David Sacks, a former Trump AI adviser and a vocal critic of Anthropic’s safety approach, warned that the US deviates from a pro-innovation, pro-export strategy “at our peril.” The administration, for its part, points to how quickly it acted; a Commerce Department spokesman said the goal was to keep America the global AI leader while safeguarding security.
The Open-Weight Problem
At the core is a governance gap the controls do not address. US restrictions assume a vendor sits between a model and its users, able to monitor and cut off abuse. Open-weight Chinese models remove that layer entirely: anyone can download GLM-5.2 or Moonshot’s Kimi K2.7 and run them on private servers, leaving no provider-side logs.
One security startup founder called it the “Wild West,” and said the models are improving at cyber tasks like analyzing reconnaissance data and generating exploit code. The shift is compounded by a wider corporate move away from unconstrained AI spending toward efficiency, which also favors cheaper Chinese options. One startup, Lindy, switched entirely to China’s DeepSeek and watched its costs collapse, while Coinbase says using open-weight models like GLM-5.2 roughly halved its AI bill.
A Policy Crossroads
The episode exposes how much harder it is to control AI models than the chips they run on. For years the US has used export controls to keep advanced Nvidia and AMD chips and Huawei gear out of adversaries’ hands, and it is now extending that logic to the models themselves, treating them as dual-use national-security assets. But early evidence suggests the restrictions may be accelerating the alternatives they aimed to slow: Chinese labs have cut prices, DeepSeek closed a roughly $7.4 billion round, and demand for Chinese models has surged on some platforms.
Allies have also chafed at their new dependence on Washington’s decisions. The unresolved question is whether the US can build a workable framework before Chinese capabilities, already near parity on some cyber tasks, close the gap entirely, leaving defenders unprepared.
