The Trump administration’s export control restrictions on the two leading US AI labs are producing the opposite of their intended effect, according to CNBC. While Anthropic’s Fable 5 has been offline for weeks and OpenAI’s GPT 5.6 series is limited to a US-only preview, Chinese model makers are shipping competitive alternatives without similar constraints.

Restrictions on Both US Frontier Labs

Anthropic received partial clearance on Friday to release its Mythos 5 model to select US companies and federal agencies, but Fable 5 remains unavailable after a two-week shutdown triggered by an export control directive. OpenAI voluntarily restricted its GPT 5.6 models (Sol, Terra, Luna) to US-only access at the government’s request on the same day, CNBC reported.

The result: both leading US model providers are operating under access constraints at the same time, while their Chinese competitors face none.

Zhipu’s GLM 5.2 Matches Frontier Benchmarks

Zhipu’s GLM 5.2, an open-weight model released earlier this month, matches top US labs on some cybersecurity benchmarks according to analysis by Semgrep. The model performs on par with Mythos in code analysis tasks, according to CNBC.

Jefferies strategist Christopher Wood told clients that GLM 5.2 “is almost equal to Anthropic as a competitor for the corporate market and is just one quarter of the cost in terms of cost per token,” citing industry sources.

Marc Andreessen called the timing “incredible given current events.” Sam Bresnick, a research fellow at Georgetown’s Center for Security and Emerging Technology, described the developments as “a pretty good wake-up call,” according to CNBC.

Enterprise Migration Already Underway

The restrictions coincide with a broader shift in enterprise AI spending from unconstrained token consumption to cost efficiency. That shift is pulling customers toward cheaper Chinese alternatives.

Flo Crivello, CEO of AI startup Lindy, told CNBC he moved 100% of Lindy’s traffic from Anthropic’s Claude models to DeepSeek. “We did it, and you could see that cost curve go down, like, crash to the ground,” Crivello said.

Travis Lanham, co-founder of AI security startup Armadin, told CNBC his team is experimenting with GLM 5.2 and Moonshot AI’s Kimi K2.7 for cybersecurity use cases, calling the open-weight model ecosystem “the Wild West.” The models are easy to download and run on private servers without relying on a third-party cloud, bypassing any access restrictions entirely.

Policy Contradiction

The irony is not lost on Washington. Former Trump crypto and AI czar David Sacks wrote on X: “A year ago, President Trump declared that America was in a global AI race and that the way to win it was to be pro-innovation, pro-infrastructure, pro-energy, and pro-export. President Trump was exactly right; we deviate from that strategy at our peril.”

For years, the US has restricted AI chip exports to China from Nvidia and AMD while banning Huawei equipment. Now a parallel dynamic is emerging on the model side: US restrictions on domestic model availability create an asymmetric window where Chinese open-weight models can gain enterprise traction while American alternatives sit behind policy gates.

The Agent Infrastructure Question

For teams running autonomous agents, the competitive dynamics are particularly acute. Agent systems require reliable, low-latency access to frontier models. When Fable 5 goes offline or GPT 5.6 is US-only, agent deployments that depend on those models either stall or switch providers. Chinese open-weight models that can run on private infrastructure without cloud API dependencies offer a workaround, but introduce their own supply chain and security considerations that most enterprises have not fully evaluated.

Anthropic, OpenAI, and the White House did not respond to requests for comment, according to CNBC.