The Bigger Risk May Be US Restrictions on Frontier Models

Kimi K3’s strong showing against Claude Fable 5 and GPT-5.6 Sol set off another round of alarm about Chinese open-weight models overtaking US labs. Ben Thompson thinks the alarm is pointed the wrong way.

What he said

Thompson’s real target is US policy toward its own frontier models, not the Chinese models everyone is watching. On cybersecurity defenders being unable to use Anthropic’s or OpenAI’s best models for security work, he writes:

“Right now defenders are effectively banned from using Fable or Sol for cybersecurity because of Trump administration directives; that means the best alternative is using models from a country which has been trying to weaken our cyber defenses for years. This is insane!”

He frames China’s approach to open-weight releases as a deliberate strategy, not charity: “The strategy for China is obvious: commoditize your complements.” Cheap, widely available AI strengthens China’s position in robotics and manufacturing, the physical-world businesses that actually benefit when software intelligence gets cheap. He also argues that restricting distillation, the practice of training a smaller model on a larger one’s outputs, from US frontier models has a side effect few intended: “Western open weight model makers must either reproduce those capabilities independently or wait to learn from Chinese models.”

Who he is

Ben Thompson is the founder and author of Stratechery, a technology strategy newsletter he has written full-time since 2014. He is an independent analyst, not an employee of any AI lab, and Stratechery’s Aggregation Theory has shaped how much of the industry talks about platform economics.

What he gets right, and where it’s incomplete

Thompson’s cost-structure argument holds up. If intelligence for routine tasks keeps getting commoditized, the labs that win are the ones that can serve it cheapest, not the ones with the flashiest benchmark score. That is a useful corrective to headlines that treat every new open-weight release as an existential threat to US labs.

His policy critique is sharper but narrower than it sounds. He is describing a specific, reported restriction on using Fable and Sol for cybersecurity defense work, not a blanket ban on all AI use by security teams. The piece does not name the exact scope of the “Trump administration directives” it references, so builders should treat that detail as Thompson’s characterization until a primary policy document confirms it.

The “commoditize your complements” framing is a strategic read, not a documented Chinese government plan. It is a plausible interpretation of the incentives at play, and worth taking seriously, but it is Thompson’s inference, not a quoted policy statement from Beijing.

Why it’s notable

Most of the commentary around Kimi K3 has been about whether it is good enough to use, which is the same ground BuilderWithin’s own open-weight coding models piece covered this week. Thompson skips that question entirely and asks a different one: what set of incentives produced this model, and what does the US response to it actually cost American builders. That is a distinct angle on the same news cycle, and it changes what a builder should actually watch.

What it means for builders

If Thompson is right about commoditization, expect API pricing for standard coding and drafting tasks to keep falling, while frontier-tier reasoning work stays priced at a premium. That is a reason to keep re-testing cheaper models against your actual workload instead of assuming your current provider is still the best price for the job.

The distillation point matters if you use open-weight coding models. Understanding that many of today’s strongest open-weight models trace back to Chinese frontier labs, partly because US labs’ terms of service block distillation from their own models, is useful context when you are choosing a model for anything security-sensitive or when data residency and provenance matter to your customers.

Thompson’s piece is a reminder that the more consequential decisions about which models you can use are often made by policy, not benchmarks. Track both.

End of article