White House Weighs a Six-Month Deadline for Open AI Models

The White House is weighing a cap on how powerful an open-weight AI model can legally be, and the people tracking the talks put the decision window at six months. No executive order has been signed naming open models directly. But the administration has already shown, twice this summer, how fast it can force a model dark.

Anthropic is the loudest voice asking for limits, and it is also the company positioned to gain the most if Chinese rivals get cut off. So far its own systems are the only ones Washington has actually taken down.

The White House Weighs a Capability Ceiling

On June 2, 2026, President Trump signed an executive order called Promoting Advanced Artificial Intelligence Innovation and Security, the administration’s biggest step yet toward federal oversight of AI. The order set up a voluntary review window letting officials test a new model for national security risk before release.

It says almost nothing about open weights. Its voluntary framework does not cover open-source models, even ones that match frontier capability.

That gap is what policy watchers say comes next. People tracking internal conversations describe a proposal to ban or indefinitely delay any open-weight model that clears the rough capability of GPT 5.5, Claude Opus 4.8 or GLM-5.2, current flagship systems from OpenAI, Anthropic and China’s Zhipu AI. No order has been made public, and no timeline is confirmed. But the six-month window keeps coming up in the same conversations.

Two Claude Models Already Went Dark for Weeks

Anthropic released Fable 5 and Mythos 5, its two most advanced systems, on a Tuesday in early June. Within days, the Commerce Department ordered the company to cut off every foreign user, after Amazon raised cybersecurity concerns about the models.

Commerce Secretary Howard Lutnick’s letter gave Anthropic roughly 90 minutes to comply, pulling access for hundreds of millions of people in one motion. Anthropic chose to pull both models entirely rather than risk partial compliance.

The blackout lasted more than two weeks. The government eventually let Anthropic resume offering Mythos to certain trusted partners, then lifted the Fable restriction completely earlier this month.

OpenAI avoided a formal order but not the pressure. It agreed to a staggered rollout of GPT-5.6, limiting its flagship Sol model, plus lower tiers Terra and Luna, to government-approved customers while Commerce Department testers reviewed it.

When Axios reported the White House had given OpenAI a green light to go fully public, officials pushed back hard.

The Trump administration did NOT give OpenAI a “green light,” approval, or clearance to release its models. No such permission is required or granted.

A White House spokesperson told Gizmodo that, adding that decisions on timing and scope of any release rest entirely with the companies themselves.

Who’s Actually Pushing for a Ban

Dario Amodei has been consistent on this point since long before Mythos existed. Anthropic’s chief executive told Congress in 2023 that scaling open-weight models further was “going down a dangerous path.”

The company has since built a specific case against Chinese labs. In February 2026 it disclosed that DeepSeek, Moonshot AI and MiniMax had together created roughly 24,000 fake accounts and pulled more than 16 million exchanges out of Claude to train rival systems. By June, Anthropic told senators that Alibaba’s Qwen team had done the same thing with about 25,000 accounts and 28.8 million interactions.

The White House’s own Office of Science and Technology Policy adopted that framing in an April 23 memo, calling the pattern “deliberate, industrial-scale campaigns” and concluding there is “nothing innovative about systematically extracting and copying the innovations of American industry.”

The fear did not start with Mythos. In January 2025, DeepSeek shocked the industry with its R1 reasoning model, proving a relatively small Chinese firm could match frontier performance on a fraction of the usual budget.

Nathan Lambert, an AI researcher whose newsletter tracks the policy fight closely, calls the wider campaign regulatory capture. His argument is straightforward: Anthropic gains real economic security if the Chinese open models it accuses of distillation get banned outright, and its public statements read more like policy recommendations than neutral evidence sharing.

That risk is not unique to this fight. A December analysis by economists Jérémie Haese and Christian Peukert found that openness is not politically neutral and can pull regulatory attention away from more structural competition concerns.

Is the Distillation Threat Overblown?

Not according to Anthropic and the White House, who call it industrial-scale intellectual property theft. Independent researchers argue the technique alone cannot manufacture a frontier model, and cybersecurity experts say a related government crackdown already went further than the evidence supported.

The hawkish case is laid out plainly by national security researchers writing in Foreign Affairs, which described industrial-scale distillation of American frontier models, compressed to run cheaply and shipped back into the United States itself.

Other analysts think that case skips a step. Distillation can copy a teacher model’s outputs, the argument goes, but not the research, data pipelines and infrastructure that built the teacher in the first place. On this view, the technique produces a fast follower, not a new frontier lab.

Zhipu AI shows how tangled this gets. The company has sat on the U.S. Entity List since January 2025, yet its technical reports reveal heavy reliance on DeepSeek’s architecture, the same lab Anthropic and OpenAI both say engaged in systematic distillation. Capabilities that started inside an American lab may have already completed a full loop through China’s open ecosystem, past a sanctions list meant to stop exactly that.

Three camps do not agree on what to do about any of it.

• Cybersecurity researchers, including Stanford’s Alex Stamos, argue the Fable and Mythos export controls had no demonstrated unique risk behind them and mainly took capable defensive tools away from security teams.
• National security writers in Foreign Affairs argue the U.S. could win the training race on raw capability and still lose what they call the distribution war if distilled Chinese models keep spreading faster.
• Ethan Mollick, a University of Pennsylvania professor, thinks a U.S.-only ban would not solve the underlying problem, since Beijing faces the identical dilemma once a model crosses into risky territory.

Stamos put his objection to Commerce Secretary Lutnick and National Cyber Director Sean Cairncross directly on a call with reporters.

“I just want to say that pretty much nobody in the cybersecurity industry believes that there’s any factual basis for this action,” Stamos said.

Where the Open Model Market Has Already Moved

The clearest evidence of where developers are actually voting sits on OpenRouter, a platform that routes requests across AI providers. Google, Anthropic and OpenAI’s combined share of usage dropped from 55% to 33% between January and June 2026, before this summer’s restrictions had even fully played out.

Z.ai, the Chinese startup also known as Zhipu AI, helped trigger that shift when GLM-5.2 landed in mid-June, scoring within a few points of top American systems on coding and agentic benchmarks. Security firm Semgrep found it outscoring Claude Code on an access-control bug benchmark at roughly seventeen cents per bug found.

“GLM-5.2 is free to download, fine-tune and run on an enterprise’s own servers, putting pricing pressure on frontier labs at the same time that access looks shaky,” AI analyst Andrew Curran noted.

The Companies with the Most to Lose

A ban would not just reshape the frontier lab race. It would land on a specific, growing list of businesses that already made their choice.

• Airbnb and Cursor are both named in the same congressional distillation investigation that identifies Moonshot AI, maker of the Kimi model, as a Chinese firm accused of coordinated distillation campaigns.
• Coinbase has already cut its spending on Chinese-made models in half, according to a separate Tech Times report citing legal exposure rather than performance concerns.
• US inference and fine-tuning startups built entire businesses on the assumption that open weights keep improving. A capability freeze cuts off their product roadmap along with everyone else’s.
• Ordinary enterprise adopters who switched to cheaper open models to escape rising per-token costs on closed systems now face the same uncertainty that just hit Anthropic’s own customers.

None of these companies asked for a place in a national security debate. They got one anyway, the moment cost pressure pushed them toward open weights.

The One Off-Ramp Left

Lambert’s own prescription is narrower than any policy fix. Get an American company to ship an open model that matches the frontier, and the entire conversation changes from only China builds capable open models through distillation to a shared problem the whole industry has to manage together.

Two firms fit that description on paper. Meta and Microsoft both benefit from wide access to AI without needing to monetize a model directly, the kind of company economists call motivated to commoditize a complement. Among Western firms today, though, France’s Mistral stands largely alone in championing open weights, while Meta, once a vocal open-source advocate, has stepped back under new leadership.

Washington is not the only capital weighing this trade-off. Most Chinese labs currently ship open-weight, and that openness is exactly how they built global reach despite trailing America’s best models by about seven months on average.

A trailing player rarely gives up its biggest advantage unless the concern driving the decision is genuinely national security, not competition.

Frequently Asked Questions

What’s the difference between open-source and open-weight AI models?

Open-source usually means the training data, code and process are all public, while open-weight only means the finished model’s parameters are released for anyone to download and run. DeepSeek’s R1, for example, shipped under the permissive MIT License with six smaller distilled versions included, a looser standard than most Chinese open releases actually meet.

Why do US labs consider distillation a threat?

Distillation lets a rival skip the most expensive part of building a frontier model. A smaller system can be trained to copy a stronger “teacher” model’s outputs instead of learning from scratch, which otherwise costs tens or hundreds of millions of dollars for a genuinely frontier system.

Would OpenAI’s own open models fall under the proposed ban?

Probably not. OpenAI’s existing open-weight releases, a pair of models called gpt-oss, are considered far less capable than its private flagship systems and sit well under the GPT 5.5 line that policy watchers say would trigger scrutiny.

Has the US government banned an AI model already?

Not an open-weight one. The only models pulled so far, Anthropic’s Fable 5 and Mythos 5, were closed systems, and the export-control order banned foreign nationals from using them whether they were inside or outside the United States, a far broader reach than a typical sanctions list covers.

Could China end up restricting its own open models?

It’s possible. Reporting on Beijing’s internal deliberations found Chinese officials are debating whether to restrict export of their own most powerful open-weight systems, even though that openness is exactly how they closed a gap that still runs roughly seven months behind America’s frontier.

Is there political pushback against federal control over AI rules?

Yes. A bipartisan coalition of 36 state attorneys general already wrote to Congress in November 2025 opposing a federal ban on state AI laws, a sign that any executive action aimed at open models could run into a similar states’ rights fight.