June 22, 2026

3 Minute Read or Listen ->

🔱 Sign up for a free demo today!

See how the latest AI and digital advocacy tools can transform your work.

THIS WEEK'S EDITION

 The Fable Shutdown Was Meant To Contain China. It's Doing The Opposite.

On June 12, the U.S. government pulled the plug on the most powerful AI America had ever shipped. The stated goal was to keep frontier capability out of foreign hands. Instead, the shutdown stampeded customers toward the one part of the market the ban can't touch — most of it Chinese. In today's short read, we explain why the shutdown's biggest winner is Beijing, and how public affairs teams should adapt.

The Barn Door Was Already Open

Anthropic launched Fable 5 and Mythos 5 on June 10. Two days later, on June 12, a federal export-control directive forced the company to shut both models down for every customer on Earth. As of this writing, they remain offline, with no restoration date in sight. But for the two days they were live, the most capable model on the planet was open to anyone with a login — including the foreign nationals the directive then attempted to lock out.

The stated rationale of the directive was national security. But the order landed amid a months-long feud with Anthropic — the Pentagon has branded the company a "supply-chain risk," and Anthropic is suing the government over what it calls retaliation. Critics call the ban punitive, noting that it doesn't apply to Anthropic rivals like OpenAI, whose models would face the same logic.

But this fight misses something even more important: China's best labs no longer need a Fable login. American frontier labs accuse the top Chinese models of "distillation" — quietly training on the outputs of U.S. models. In February, Anthropic said DeepSeek, MiniMax, and Moonshot ran industrial-scale campaigns — roughly 24,000 fake accounts and 16 million interactions — to extract Claude's capabilities. Does anyone seriously think Beijing waited politely for Anthropic and the Trump Administration to resolve their standoff before distilling Fable?

China's Open Models Just Got A Gift

Cut off a closed model, and customers hunt for one that nobody can switch off: enter open-weight models you download and run yourself — like owning a copy of the brain instead of renting access to it. 

The best open models are now Chinese. On June 13, one day after the Fable shutdown, Z.ai released GLM-5.2, a free, MIT-licensed model whose self-reported coding scores land just behind Claude Opus 4.8 — Anthropic's most capable publicly available model — while requiring a fraction of the computing power. DeepSeek's V4 has effectively closed the coding gap with American frontier models, scoring within a fraction of a point of Claude on the industry-standard SWE-bench Verified test. On OpenRouter, a major model marketplace, Chinese models now hold the top four spots. The Chinese labs aren't subtle about it — Z.ai launched under the tagline "frontier intelligence belongs to everyone."

And the cost gap is just as big a story. Claude Opus 4.8 runs about $25 per million output tokens; DeepSeek V4 charges roughly $3.48, and its lighter Flash tier just $0.28. GLM-5.2 sits in the same bargain range. By one analysis, a workload of 10 million output tokens a day costs about $260 a month on DeepSeek against roughly $7,500 on Claude.

Even Microsoft Is Shopping In Beijing

Microsoft has now confirmed it's evaluating a self-hosted, fine-tuned version of China's DeepSeek as a cheaper engine for Copilot Cowork, its agentic assistant in Microsoft 365, with a decision due within weeks. Read that again. An American tech giant may build part of its flagship AI product on a model from the country we just rewrote export law to keep out. And it doesn't stop there: by one congressional estimate, 80% of U.S. AI startups now use Chinese open models.

What This Means For Public Affairs

Open models are no longer a fringe option you can wave off on security grounds. The worry that spooks every general counsel is data leaving the building for servers in China. Self-hosting fixes that, and there are two paths. You can run an open-weight model on your own hardware, and the data never leaves the building. Or rent U.S. cloud compute, no hardware required: AWS already offers DeepSeek through Bedrock, in a private setup where your data stays on American servers and never reaches China. That second path is Microsoft's route, on its own Azure cloud — and it's open to organizations of all sizes.

The U.S. has had its best player benched for 10 days now, just as China was already catching up. For most organizations, the result is clear: Chinese open models are no longer something you can afford to ignore.

Two Ways To Operationalize This Email

🔱 Audit your AI dependencies. List every tool your team relies on that runs on a single closed model. Consider what open models are roughly equivalent for your uses.

🔱 Run a self-hosted pilot. Stand up an open-weight model — on your own hardware or rented U.S. cloud like AWS. Check the results and the cost, then reconsider your tech stack.

🔱 If you're enjoying this content, please consider forwarding this email to a colleague or friend. 

🔱 If you're not already a subscriber, please sign up here to stay up to date on the latest developments in political technology.

🔱 If you have suggestions for future newsletters or any other feedback, we'd love to hear from you.