The Trump administration and AI industry are discussing a capability framework that would tie U.S. open-source model releases to the capabilities of leading Chinese open-source models, with experts predicting a potential ban on frontier open-weight models within six months as Mythos-class cyber capabilities become freely downloadable.
The Trump administration and AI industry groups are discussing a capability framework that would streamline U.S. open-source model releases only if their capabilities match or fall below those of leading Chinese open-source models, a person close to the discussions said — a policy that could effectively cap open-weight AI development at current Chinese frontier levels.
"The most likely incoming action is to ban or indefinitely delay any open-weights model meaningfully above the capability level in the range of GPT-5.5, Claude Opus 4.8, or GLM-5.2," Nathan Lambert, a machine learning engineer formerly at Meta and Hugging Face, wrote in his widely-read newsletter Interconnects. "With the consistent capability gap, this should be within the next six months."
The conversations center on clarifying guidance for the AI executive order the administration issued in June, not a separate forthcoming order as online speculation has suggested, the person said. The proposed mechanism would require the Commerce Department's Center for AI Standards and Innovation to develop a procedure for comparing U.S. open models against Chinese ones. Chinese Mythos-class models — AI systems with cyber capabilities comparable to Anthropic's most advanced offering — are expected to be available for free download within six to 12 months, according to industry estimates.
The national security calculus behind the framework
The policy push reflects growing concern that Chinese open-source models could act as "Trojan horses" for malicious software, with developers potentially leaving back doors exploitable by the Chinese Communist Party, said Daniel Remler, a senior fellow at the Center for a New American Security. "These models have documented serious security vulnerabilities that make them more susceptible to adversarial attack," Remler said, citing research into backdoor sleeper agents in Chinese AI models.
Lawmakers have escalated scrutiny. The House Select Committee on China recently announced a joint investigation into Airbnb and Anysphere over their use of Chinese AI models, citing national security risks. The concern is not limited to government use: businesses across the globe that build on Chinese open-source models may be exposing themselves to adversarial compromise.
Unlike closed models — which the government can force companies to retract — open-source models, by definition, cannot be recalled once published. "With open source, they can try and sort of train safeguards into the model, but it's very easy to train them out again," said Helen Toner, executive director at Georgetown's Center for Security and Emerging Technology and a former OpenAI board member. "So I do think that should affect the risk calculus."
Distillation debate and the Anthropic factor
The capability framework discussions unfold alongside a parallel battle over model distillation — the practice of using one AI model's outputs to train another. Anthropic CEO Dario Amodei has warned that Chinese companies are systematically distilling capabilities from U.S. frontier models, with U.S. officials estimating unauthorized distillation costs American AI labs up to $6 billion annually in lost revenue.
"What I do worry about with some of these laggard models is the risks of them, where we have Mythos-class cyber capabilities," Amodei said in a June Bloomberg interview. "Months from now, Mythos-class cyber capabilities may just be available for anyone to download."
Lambert argues the distillation campaign has become a form of regulatory capture. "The action that Anthropic is effectively asking for is the wholesale banning of pretty much all the Chinese open weight models in the U.S.," he wrote. "This would demolish the open model economy that is emerging in the U.S. with inference companies, fine tuning companies, new products, and everything in between."
The data so far suggests frontier labs are not yet feeling competitive pressure. On Vercel's AI Gateway, DeepSeek's V4 Flash processes about 5.3 trillion tokens per week — more than double the 2 trillion tokens handled by Anthropic's Opus 4.8. Yet Opus 4.8 costs about $1.37 per million tokens versus V4 Flash's $0.06, a 23x premium, meaning Anthropic likely still captures the majority of platform revenue despite lower usage volume.
What comes next
A capability-based framework would not change the fundamental timeline. If Chinese Mythos-class models become freely downloadable within a year, the U.S. faces a choice: allow domestic open-source models at equivalent capability, or attempt to ban Chinese models outright. A senior GOP aide working on AI policy said a ban on Chinese open-source models would face practical difficulties but that "that's not to say Congress won't try."
Open-source advocates argue the only viable path is for a U.S. company to release a comparably capable open model, shifting the narrative from "only China builds open frontier models" to a shared ecosystem. Microsoft and Meta, which have business incentives to commoditize AI infrastructure, are seen as the most likely candidates. Reflection AI, which has argued for capability-based exemptions in meetings with the administration, has not yet released a public model.
"I don't think there's a straightforward line from, 'Open-source models will reach Mythos-level capabilities,' and then they will be permanently banned," Toner said. "The online security environment will have changed also, and we will have had several months to kind of get used to having this level of capability."
This article is for informational purposes only and does not constitute investment advice.
