Washington tried to keep Anthropic's strongest models under tighter control. In the first two weeks, it gave Asian AI labs a sales pitch they couldn't have bought.
The useful question isn't whether the Trump administration had a real cyber risk to worry about when it moved against Anthropic on June 12, 2026. It did. The question is whether cutting off access to Claude Mythos 5 and Fable 5 made that risk smaller, or simply told every serious lab outside US jurisdiction exactly what to build next.
Axios reported that Commerce Secretary Howard Lutnick's June 12 order effectively forced Anthropic to take its most powerful models down while the government reviewed diversion risks and cybersecurity safeguards. By June 26, the department had partly relented. Mythos 5 could come back for named entities in an approved annex, including their foreign national employees, and for Anthropic's own foreign national staff. Fable 5 stayed restricted.
That is not a clean reversal. It's a warning label on dependence.
Tokyo-based Sakana AI moved fast into that opening. Its Fugu technical report, posted on arXiv on June 19, describes Fugu and Fugu-Ultra as orchestrator models that route work across teams of AI agents rather than trying to be one all-conquering model. That detail matters. Sakana didn't have to beat Anthropic model for model. It had to show developers in Seoul, Singapore, Mumbai or Berlin that a single endpoint could keep working even when Washington changed the rules.
Fugu's appeal is brutally practical. A company building on one frontier provider is making a technical decision and a political one at the same time. If the model can disappear because an export-control letter lands on a Friday, your architecture has a weak point that has nothing to do with latency, benchmark scores or developer experience. Sakana saw that and sold around it.
As Tom's Guide noted after the release, Fugu acts less like a standalone chatbot and more like a manager of specialized models, selecting systems for parts of a task, checking their outputs and assembling the answer. The Sakana paper says Fugu-Ultra is designed to prioritize answer quality on difficult problems, including software engineering, reasoning and scientific benchmarks such as SWE-Bench Pro and GPQA-Diamond. You don't need to treat every benchmark claim as gospel to see the point. The product is built for a world where access is unstable.
China's 360 Security Technology made the same point in a louder register. At ISC.AI 2026 at the Beijing National Convention Center on June 24, founder Zhou Hongyi introduced two tools under the Yitian Tulong name, according to TechRadar's write-up of Reuters reporting. Tulongfeng is aimed at vulnerability discovery. Yitianzhen is built for cyber defense and incident response. Zhou called Tulongfeng a Chinese Mythos.
He also didn't pretend China had already caught up. Zhou said domestic models still lag US models by 20% to 30% in base capability, but argued China couldn't wait until that gap closed before automating vulnerability work. 360 claimed Tulongfeng had found 3,432 software flaws, including 105 confirmed by the Chinese government. TechRadar noted that those claims could not be independently verified.
Keep that caveat in the story, because it is the story. Zhou doesn't need you to believe every number to understand the pitch. He is telling Chinese companies and officials that the right answer to restricted US models is not patience. It is domestic tooling, proprietary security data and automation wrapped around models that may still be weaker at the core.
Z.ai gives the same trend a market face. The company, formerly known internationally as Zhipu AI, has spent the past year pushing GLM models into the open and building a public-company profile in Hong Kong. The South China Morning Post reported in April that Z.ai open-sourced its flagship GLM-5.1 model while raising API prices by 10%, and its shares rose that day. That is the kind of company that benefits when customers start asking whether US model access is a permanent entitlement or a revocable license.
For investors, Sakana is the cleaner signal than 360. Fugu isn't a sovereign model in the old sense. It doesn't magically free users from US, Chinese or European infrastructure politics, and it still depends on the quality and availability of the models it orchestrates. But it sells resilience as a product. Frankly, that may be easier to monetize than national AI pride.
The US concern around Mythos 5 is not nonsense. A cybersecurity model that can help find and exploit software vulnerabilities at scale deserves scrutiny, especially when critical infrastructure is involved. But policy has effects beyond the memo. In this case, the first visible effect was to make revocability a boardroom problem and a product category.
Washington wanted to slow the spread of dangerous capability. In the short run, it may have done something narrower and stranger: it taught the market to pay for AI systems that route around Washington.
Also read: Apple is asking Washington to let it buy chips from a blacklisted Chinese firm and the answer will matter far beyond Apple, Steve Eisman says investors betting on hyperscalers in the AI race are funding the wrong side of the trade, and Satya Nadella says companies that rent their AI brains are making a strategic mistake they will regret
