18 days.

That's how long US government export controls on Anthropic's strongest model, Mythos 5, lasted. On June 12, citing "national security authorities," a control order was issued and Anthropic was forced to shut down global public access; on June 26, the Commerce Department notified that access could be opened to "White House-approved customers"; on July 1, restrictions were fully lifted and global access was restored.

Commerce Secretary Howard Lutnick's conditions for lifting controls: Anthropic agreed to "proactively detect and respond to model-related safety risks, cooperate with the US government on future model release protocols, and promptly report malicious activity." But Anthropic had already proactively made these commitments months ago through Dario Amodei's public policy writings. In other words, the conditions the government set for lifting controls were things Anthropic was already doing.

Why They Caved After 18 Days

The core reason controls were lifted after just 18 days is simple: the blockade didn't work.

During those 18 days Mythos was blocked, Asian AI companies' models were rapidly closing in. Japan's Sakana AI released the Fugu model; China's Z.ai (Zhipu) released open-weight GLM-5.2 -- The Verge reported that some researchers claimed GLM-5.2 could already match Mythos levels in certain vulnerability discovery and cybersecurity scenarios. While GLM still trails Anthropic and OpenAI models on general tasks, it's an undeniable fact that "China has significantly narrowed the gap."

This is the fundamental paradox of export controls: you block the strongest model thinking you'll maintain technological superiority, but open-source models and international competitors quickly fill the vacuum, and your blockade instead cedes the global market to rivals. Cybersecurity experts were skeptical of these controls from the start -- in their view, this looked more like political pressure (Anthropic executives had publicly criticized the government's approach to AI technology) than genuine security measures.

Anthropic reached an agreement with California Governor Newsom for the state government to use Claude at half price -- further illustrating that while the federal government targets Anthropic, state governments are actively embracing AI. The chaos and fragmentation of US AI policy is plain to see.

AI Policy Enters a "Wobble Period"

Mythos's 18-day restriction-and-lift isn't an isolated case; it exposes deep contradictions in the Trump administration's AI policy.

On one hand, the government remains wary of frontier models -- OpenAI's GPT-5.6 Sol was also required to have a "controlled release," not open to all users but first to "approved organizations"; an executive order in June required pre-release review of models. On the other hand, the US can't truly block its most competitive technology -- AI is America's core chip in the tech competition with China, and handicapping its own industry doesn't serve strategic interests.

This whiplash creates enormous uncertainty for the entire industry. Dean W. Ball (an analyst who recently joined OpenAI's policy team) criticized the June executive order, arguing the review mechanism is unclear. Companies don't know if their next model release will get blocked; investors don't know whether to apply a "policy discount" to valuations; international customers don't know whether to trust American suppliers or nurture domestic alternatives.

The more far-reaching impact: the "18-day fiasco" of export controls sent a signal to the world -- American AI supply is unreliable. This will accelerate determination across countries (especially the EU, China, and the Middle East) to develop domestic AI capabilities, and will make international customers more inclined toward "multi-vendor strategies" rather than deep lock-in with American companies.

Mythos is unblocked, but the Pandora's box of AI geopolitics has been opened. Who will be the next to face controls? How long will restrictions last? Nobody knows. The only certainty: in AI, technology iterates faster than policy can be written, and trying to constrain a technology that updates monthly with traditional export control frameworks is a losing race from the start.

See you tomorrow.