For two weeks the story was about Anthropic. Now it is about the whole American frontier. On Friday OpenAI said it would limit the release of its newest models, the GPT-5.6 family known internally as Sol, Terra and Luna, opening them only to a small group of trusted partners whose names have been shared with the government. The same day, Anthropic was finally cleared to offer its Mythos 5 model to "certain trusted partners," though its Fable 5 model remains off the market. Two of the country's leading labs are now waiting for permission to ship their best work.
The mechanism is new and still half-built. Under an AI security executive order President Trump signed earlier this month, developers are asked to voluntarily share models with advanced hacking capabilities with the government for up to 30 days, so the Commerce Department's Center for AI Standards and Innovation can assess the risks to financial systems and national security. Notably, OpenAI is complying with a request, not a binding order. Chief executive Sam Altman reportedly told staff the administration would approve access "customer by customer" during the preview. The company was pointed about its discomfort: "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."
What gives the moment its edge is the timing. The original logic of America's hands-off stance was that any drag on domestic AI would hand the lead to China. That fear is now being tested in real time. Researchers say Zhipu's GLM-5.2, released earlier this month, performs on par with top US labs on some cybersecurity benchmarks, in places matching Mythos itself. A Jefferies strategist told clients it 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." The reaction from Silicon Valley's loudest voices was not triumph but alarm. Marc Andreessen called GLM-5.2 the first Chinese model to "match and often beat" the American labs "with no compromises," adding, "Incredible timing given current events."
The competitive problem is sharpened by how these Chinese models travel. Many are open-weight, meaning a company can download them and run them on its own servers, no cloud contract or export clearance required. Citi noted this week that the intelligence gap between proprietary and open-weight models has compressed sharply, with GLM-5.2 a driver, just as regulators slow the release of the closed frontier. The result is a quiet migration. The startup Lindy moved all of its traffic to China's DeepSeek and watched its cost curve "crash to the ground." Coinbase says it now leans on GLM-5.2 and Moonshot's Kimi to cut nearly half its AI spending. Shopify and Airbnb have praised Alibaba's Qwen 3.
This is the awkward shape of the policy. The reviews are a genuine attempt to keep tools that can automate cyberattacks out of the wrong hands, and few dispute that the underlying risk is real. But a pause that only binds American labs, while open-weight Chinese models spread freely and undercut them on price, risks achieving the opposite of its goal. Even David Sacks, the former AI czar and a frequent critic of Anthropic, warned against deviating from a "pro-innovation, pro-export" course. Elon Musk predicted GLM would reach Fable-class capability by the first quarter; Zhipu's founder replied that it "won't take that long." Washington wanted a 30-day look before the most powerful models ship. The question is whether 30 days is long enough for the lead it is trying to protect to slip away.