The US AI Crackdown Handed Chinese Open-Source Models a Boost

The Trump administration in June ordered Anthropic to cut off foreign users from its Mythos 5 and Fable 5 frontier models, then put OpenAI’s flagship GPT-5.6 behind a US per-customer approval gate. Builders moved. The Anthropic shutdown ended after the Commerce Department lifted the controls on June 30, with full global access for Fable 5 restored on July 1. The usage did not return. On OpenRouter, the world’s largest AI model routing platform, the combined share of Google, Anthropic and OpenAI fell from 55 percent in January to 33 percent in June, and China’s open-weight DeepSeek now leads outright, in an account of the US crackdown on frontier AI.

The gap was filled by Chinese open-weight systems, Z.ai’s GLM-5.2 among them, plus DeepSeek, Moonshot AI’s Kimi K2.5 and Alibaba’s Qwen team. “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 said. The pattern has held even after Washington eased the rules.

The Crackdown That Started the Flight

In early June, the White House signed an executive order titled “Promoting Advanced Artificial Intelligence Innovation and Security” that asked AI developers to voluntarily submit new models for federal review ahead of any public release. The order was light on specific details. It gave agencies 60 days to write the framework. Within ten days, the Commerce Department was already exercising the power on Anthropic’s two top models.

On June 12, the Commerce Department sent Anthropic an export control directive suspending access to Mythos 5 and Fable 5 for every foreign national. Anthropic had no workable way to screen users without locking out its own foreign staff, so it pulled both models offline entirely on June 13. A few days later, OpenAI agreed to limit GPT-5.6 to customers pre-approved by the US government.

What Got Pulled Offline

Fable 5 had launched on June 9, the most advanced public Anthropic release to date, and Mythos 5 had shipped inside the same launch window. Three days after the public debut, the export order arrived and both closed to every user everywhere.

Anthropic’s three-week shutdown

Anthropic stayed dark across both products for nearly three weeks. Commerce Secretary Howard Lutnick told the company on June 26 that the government was prepared to restore access for a select list of “trusted partners.” Mythos 5 came back online for some US organizations the same day, under Anthropic’s Glasswing defensive-security program. Fable 5 returned to all global users on July 1, the day after the Commerce Department confirmed that the export controls had been lifted.

GPT-5.6 behind a US approval gate

On June 26 OpenAI launched GPT-5.6 in three variants: Sol, its flagship, plus Terra for daily work and Luna, a smaller and cheaper option, with every user pre-vetted by the US government as part of a limited preview of GPT-5.6. OpenAI itself pushed back on the regime. “We don’t believe this kind of government access process should become the long-term default,” the company wrote, adding that “It keeps the best tools from users, developers, enterprises, cyber defenders, and global partners who need them.”

OpenAI framed the rollout as a temporary preview, not the long-term shape of its business. The decision was tied to the same June executive order that had triggered the Anthropic action. As of late June, GPT-5.6 was still running behind the gate, with a broader release lined up for the weeks that followed. The arrangement forced every buyer of the latest OpenAI technology to wait in line behind a federal review. For builders who had already been through three weeks of a darkened Anthropic, the second US vendor on a leash was the moment many of them looked elsewhere.

Where the Usage Actually Went

OpenRouter, the world’s largest AI model routing platform, is where developers and enterprises pick a model for every query, and its usage split across providers is the cleanest read on what builders actually chose. DeepSeek now leads the leaderboard outright, by a clear margin. The redistribution of share during the first half of 2026 is the closest thing to a verdict on which side of the open/closed line held up.

The pivot is large enough that JPMorgan Asset Management strategist Michael Cembalest made it the centerpiece of his Eye on the Market commentary this spring, citing the data in his Memorial Day podcast. His read of the OpenRouter data lines up with the redistribution below, which captures the 61 percent Chinese share among the OpenRouter top 10 in three rows.

The Anthropic shutdown set off the redistribution but did not cause it on its own. The deeper driver is that Chinese labs had been posting frontier-tier open weights at steadily cheaper prices, a pricing trajectory the US labs have not matched. The US controls landed on a market already leaning that way, and gave the lean a sharp shove.

Why Chinese Open Models Are Winning

Three things matter to a builder picking a model: price, performance and whether it is there tomorrow. Andrew Curran, the AI analyst, framed Z.ai’s GLM-5.2 release that way on the day it shipped, June 13. The model’s weights were free to download, fine-tune, and run on any company’s own servers, and pricing pressure on US frontier labs is the through-line. Once the weights are posted, no regulator, no founder and no foreign government can pull them back. That mechanic is what pushed US frontier labs from US-default to one-option-among-many in enterprise procurement.

OpenRouter’s 100-trillion-token State of AI study, published earlier this year, named the dynamic the “Glass Slipper” effect: developers who pick a model early and tune workflows around it tend to stay. OpenRouter COO Chris Clark said Chinese open-weight models have captured so much share because they are “disproportionately heavy in agentic flows run by U.S. developers.” Programming is now the dominant share of OpenRouter usage, expanding from roughly 11 percent of total volume in early 2025 to over 50 percent by 2026. In other words, the audience that switched over was the most expensive one to lose: the build pipelines for the next round of AI products.

• June 13, 2026: Z.ai ships GLM-5.2 as an open-weight model
• June 26, 2026: OpenAI rolls out GPT-5.6 in three variants, each gated by US government approval
• June 30, 2026: US Commerce lifts export controls on Anthropic’s Mythos 5 and Fable 5
• 50%+: OpenRouter share of token volume now driven by programming-related workflows, up from about 11% in early 2025
• Free vs. revocable: GLM-5.2 weights are downloadable and self-hostable; closed US frontier models can be switched off by a single government rule

The Western Side of the Open Bet

The open-source side of the AI map is almost entirely Chinese now, with one French exception. Among Western companies, France’s Mistral “stands largely alone in championing open models,” per the Economic Times. US tech giant Meta, once the loudest Western voice for open AI, has stepped back from that line.

Meta replaced its open-source Llama family with a closed commercial model, Muse Spark, in April. The shift left the open-weight camp thinner on the Western side at the exact moment US policy was propelling builders toward it.

• Anthropic disabled Mythos 5 and Fable 5 entirely on a US export control order in mid-June
• OpenAI put GPT-5.6 behind a US per-customer approval step on June 26
• Meta closed its flagship model family with Muse Spark in April
• Mistral in France remained the lone Western company still championing open AI

For a buyer choosing between a US frontier lab and any other option, the calculus moved sharply this quarter. Closed-side reliability became a working concern, not an abstract one. An enterprise that had picked Anthropic’s Fable 5 on June 9 had no Fable 5 by June 13. An enterprise that picked GPT-5.6 on June 26 was on the US government’s contact list by launch day. Meanwhile, a developer who downloaded GLM-5.2 on June 13 still has GLM-5.2, on hardware the developer controls.

What Builders Did When the Doors Closed

Smaller builders felt it hardest. Haitham Mengad, co-founder of Stems Labs, an AI music startup, had built production work around Anthropic’s Fable 5. When the company pulled the model on June 13, he called it the first time he saw how brittle the closed-AI bet could be. “Fable has been a game-changing model for me. Honestly, when they took it off, it was the first time that I realized… it’s almost like a drug,” he said.

You want to be as flexible as you can be. Maybe a year and a half ago some large company might say we bought Anthropic or we bought OpenAI, and now no one, no one buys only one.

The line is Oren Michels’s, CEO of Barndoor AI, in the Economic Times. It is the line procurement officers have been quoting inside their teams since mid-June. The shift it names is the same shift the OpenRouter share chart already shows in numbers.

The Catch No One Is Naming

The same logic that ended Fable 5 could end GLM-5.2 someday. Ethan Mollick, a professor at the University of Pennsylvania and a leading voice on AI, made the point on the record: “If Mythos-level models are considered risky, China will also not want them to be open.”

That future risk is the open-source community’s open question. For now, the practical effect runs the other way: the harder Washington pushes on US frontier labs, the more enterprise workflows migrate toward open weights that no single government can revoke. The OpenRouter data is the clearest evidence yet that builders behave in line with that logic. In other words, the policy that started the flight also keeps it from reversing.

OpenAI itself has conceded the point. In the rollout note for GPT-5.6, the company said government-by-government control of model access “keeps the best tools from users, developers, enterprises, cyber defenders, and global partners who need them,” and warned against making it the default. The complaint reads differently now that the alternative on the receiving end is a Chinese open-weight model Washington cannot pull back down.

A year ago, the playbook said buy one of the big three and integrate it tightly. The Trump administration’s AI order, and the Anthropic and OpenAI actions that followed it, made that playbook visibly fragile in June. Builders who lived through three weeks without Fable 5 will not forget the playbook breaking. That same playbook, on the other end of the network, sent more code paths through GLM-5.2, DeepSeek and Kimi K2.5 in a single month than in all of 2025. The next phase of the contest is being run in parallel across the supply chain, with Chinese influence operations already probing US AI policy debates even after the export controls came down on June 30.