Two weeks after the Trump administration banned Anthropic's most powerful models from non-US users, Asian AI labs are racing to fill the void, and the early results suggest Washington may have accelerated the competition it was trying to prevent.
On June 12, 2026, the US government ordered Anthropic to cut off foreign nationals from Fable 5 and Mythos Preview on national security grounds, citing fears that Mythos-class models could meaningfully accelerate sophisticated cyberattacks against critical infrastructure. Anthropic complied by disabling both models globally while it sorted out access controls. The pause lasted days. The competitive response it triggered may last years.
Tokyo-based Sakana AI launched Fugu on June 22, ten days after the ban took effect. The pitch was direct: "frontier capability without the risk of export controls." Sakana says its Fugu Ultra system scores 73.7% on SWE-Bench Pro, ahead of GPT-5.5's 58.6% and Gemini 3.1 Pro's 54.2%, and posts 95.5 on GPQA-Diamond against Mythos Preview's 94.6. The firm isn't claiming it built a better model. Fugu Ultra is an orchestration system, a meta-layer that routes prompts across GPT, Gemini, Claude's available tiers, and other models, verifies their outputs, and recurses until satisfied. The single API endpoint means a developer in Seoul or Singapore gets something that benchmarks like a frontier model without touching a model Washington can revoke.
That architecture is the whole story. Sakana didn't outcompete Anthropic on raw capability. It designed around the revocability problem, and it went to market with that design while Fable 5 was still offline for most of the world. The timing wasn't coincidence: as VentureBeat reported, the export ban became the product's clearest selling point before the launch deck was even public.
If Sakana's move was surgical, China's 360 Security came in louder. At ISC.AI 2026 in Beijing on June 25, founder Zhou Hongyi introduced a pair of tools grouped under the name Yitian Tulong, drawn from a well-known Chinese martial arts story about legendary weapons. The first, Tulongfeng, targets automated vulnerability discovery. The second, Yitianzhen, focuses on cyber defense and incident response. Zhou called it "China's version of Mythos."
He was also unusually candid about the gap. "Objectively speaking, domestic models still have a 20% to 30% gap in base capability," he said, according to reporting by Reuters. His bet is that layering AI models over proprietary security knowledge, vulnerability datasets, and automation pipelines closes enough of that gap for practical purposes. 360 claims Tulongfeng has already found 3,432 software vulnerabilities, with 105 confirmed by Chinese authorities, though Reuters said it could not independently verify the figures. Whether or not the claims hold up, the framing is pointed: the export ban handed Zhou the exact marketing angle he needed, and he used it.
Sakana and 360 aren't operating in a vacuum. Zhipu's GLM-5.2 topped BridgeBench Reasoning within 30 hours of the ban announcement, beating Fable 5 on that leaderboard. DeepSeek closed a record funding round of roughly $7.4 billion. Shares in Z.ai jumped more than 30% after it released a new open-source model. The order became free advertising for any lab outside US jurisdiction offering models that "can't be revoked," and the market responded accordingly.
There's a genuine question here about whether the ban was ever designed to hold. The US government lifted the block on Mythos 5 this week, clearing it for use by more than 100 US agencies and companies, but the non-US access restrictions remain. What the episode proved is that a two-week window is long enough for a Tokyo startup to ship a product framed explicitly around the revocability problem, for a Beijing firm to stage a press event with a named rival, and for Chinese labs to capture benchmark headlines and funding momentum. The competitive landscape that existed on June 11 no longer exists.
For investors, the Sakana case is probably the more interesting signal. Fugu isn't a sovereign model in the traditional sense: it doesn't run on domestic chips or train on domestic data, and it depends on the same underlying frontier models it claims to route around. But the investment thesis it represents is new. A lab that can build orchestration infrastructure specifically designed to make export controls irrelevant is selling a different kind of resilience than raw model capability, and that resilience has a price floor: whatever enterprises will pay to avoid the risk of a White House order pulling their AI stack out from under them on a Tuesday afternoon in June.
Washington's original logic for the Mythos ban was that the model could accelerate cyberattacks. That concern may be legitimate. But the actual effect of the order, at least in its first two weeks, was to surface and fund exactly the Asian AI development it was meant to constrain. Zhou Hongyi basically said the quiet part: he conceded the capability gap and launched anyway, because the political moment was too good to waste. That's not a sign of a containment policy working. It's a sign of one creating urgency on the other side.
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 • Satya Nadella says companies that rent their AI brains are making a strategic mistake they will regret