Washington Wants Chinese AI Gone. The Weights Are Already on Disk.

On June 12, at 5:21pm ET, the Commerce Department's Bureau of Industry and Security ordered Anthropic to cut off every foreign national from Claude Fable 5 and Mythos 5, worldwide, over a reported jailbreak of Fable 5's guardrails. Anthropic complied within hours. Three weeks later, on June 30, the order was lifted. In between, by multiple industry estimates, Chinese AI token volume spiked to roughly 25 trillion tokens in a single week, about 78% above US volumes for the same stretch. That is the experiment nobody meant to run, and it answered a question this month's new push to restrict Chinese AI models in corporate America has not reckoned with: which models can a government actually switch off?

A vendor is a switch. A weights file is not.

Closed, API-gated models have exactly one thing open-weight models don't: a control point. Anthropic could comply with the Commerce order in an afternoon because Claude runs behind an API Anthropic operates. Flip the access-control list, done. That is also why the order worked so cleanly as a demonstration of exactly what makes closed models exposed to exactly this kind of policy risk, and why Alex Stamos, chief product officer at Corridor, called it "a humongous own goal against the U.S. AI industry" on a Center for Democracy & Technology press call. Users didn't wait for the order to be reversed. They routed around it immediately.

Now compare that to a model whose weights already sit on a few million hard drives via Hugging Face. There is no equivalent switch. Nobody can retroactively un-download DeepSeek V4, GLM-5.2, or Kimi K2.7 Code from the machines that already have them, and nobody can stop the next copy either, short of blocking the download link after the fact, which does nothing for the copies already out.

The ban that's actually being drafted

The current round of policy attention, per reporting this month, centers on a House investigation into how much enterprise API traffic now runs through Chinese-built models, with industry estimates ranging as high as roughly 61% of tokens on some neutral routing platforms as of May. Whatever the precise figure, the direction is not in dispute: Chinese open-weight models have gone from a curiosity to a default in parts of corporate America, and lawmakers are asking whether that should be allowed to continue.

None of that infrastructure disappears if Congress passes a rule tomorrow. A ban on *new* enterprise contracts with Chinese labs is legally straightforward. A ban on *weights already running inside a company's own gateway* is a different problem entirely, closer to trying to un-ring a bell than to revoking a license.

What this means for evaluating models, not just regulating them

We built the arena to answer one question, blind: given the same one-shot prompt, what does this model actually produce? Origin, licensing, and export status don't enter into the vote, because a reader comparing outputs doesn't care whether a model was trained in Hangzhou or Menlo Park, they care whether the code runs. That's a deliberately narrow lens, and it's exactly the lens this policy fight keeps skipping past. We've run Kimi against Claude, GLM against DeepSeek, Qwen against Claude and against DeepSeek, and MiniMax against Qwen on identical prompts, and the honest summary is that the open-weight Chinese models are not winning votes on charity, they're winning them on merit at a fraction of the price. That's a separate fact from whether they should be allowed to run inside a US bank's infrastructure, but it's the fact regulators keep discovering after the policy is already written.

This was a humongous own goal against the U.S. AI industry.

There's also a straightforward asymmetry Washington keeps re-creating without seeming to intend to. Every time a closed US lab gets gated, whether by export control or by the staggered government-approval rollout OpenAI's GPT-5.6 went through in late June, the open-weight alternative that cannot be gated becomes relatively more attractive by comparison. We covered the Fable 5 suspension and reversal in Fable 5 is back; the pattern generalizes past that one incident.

• June 2: executive order requires frontier labs to disclose models to government 30 days pre-release, creates a 'protected frontier model' designation.
• June 12: Commerce orders Anthropic to block all foreign nationals from Fable 5 and Mythos 5; Opus 4.8 unaffected.
• June 12-30: Chinese model token volume estimated to spike roughly 78% above US volumes during the suspension window.
• June 30: export controls lifted, Fable 5 and Mythos 5 restored worldwide.
• July 2026: House investigation opens into Chinese models' share of US enterprise API traffic, as Coinbase, Snowflake, and Lindy publicly confirm defaulting to GLM-5.2 and Kimi K2.7 Code for cost reasons.

The honest read

A ban on new Chinese-vendor contracts is enforceable and, depending on where you sit, defensible on national-security grounds. A ban that assumes the weights can be clawed back once they're inside a company's own infrastructure is not enforceable, and the June export-control episode against a domestic company proved it by accident: the thing that made Anthropic's models bannable overnight is precisely the thing open-weight models don't have. Any policy that doesn't start from that distinction is going to keep re-learning it the hard way, one surge in Chinese token volume at a time.

European and Polish readers watching the same fight from the other side of an ocean and a different regulatory regime can follow the daily version of this beat, in Polish, at nowosci.ai. For the model-versus-model receipts referenced above, open the coding arena and run the comparisons yourself; we'd rather you see the outputs than take our word for the votes.