Tokyo's Sakana AI released Fugu on June 22 and Beijing's Qihoo 360 unveiled Tulongfeng two days later, both positioned for buyers cut off from Anthropic's most capable models.
Two Asian AI labs have shipped models in the past week that don't try to clone Anthropic's most capable systems. Instead, they route around them. Tokyo-based Sakana AI released Fugu on June 22, followed two days later by Beijing-based Qihoo 360's Tulongfeng, a cybersecurity tool the company is positioning against Anthropic's Mythos. The architecture underneath both launches is orchestration: a single model that decides which other AI system should answer a given request, including, in Fugu's case, itself.
Sakana explicitly denies the timing is political. A spokesperson told TechCrunch the Fugu release was "entirely coincidental" with the U.S. directive and that "U.S. models remain important to Asia." Co-founder Ren Ito has repeated the posture at the G7 Evian summit and in a Project Syndicate op-ed, framing the read as cooperation on access rather than decoupling.
On June 13, the U.S. government ordered Anthropic to suspend foreign-national access to Fable 5 and Mythos 5, the company's two most capable models, anywhere in the world. Anthropic's own statement frames the directive as driven by a non-universal jailbreak that surfaces minor vulnerabilities, and says access to the company's other models is unaffected. For any company, university, or government buyer outside the United States that had been planning to procure Mythos or Fable, the practical effect is the same: those models are off the table for the foreseeable future.
What Sakana built is a different kind of product. Rather than training one model to do everything, Fugu trains a language model to call other language models, including Anthropic's, Google's, Meta's, or its own, through a single API. The router picks the right one for each task and stitches the results together. The underlying research was presented at ICLR this spring. Fugu can call itself recursively, so a single request can fan out into a tree of sub-tasks handled by different systems before the answers come back. Sakana's pitch, on its Fugu release page, is direct: "delivering frontier capability without the risk of export controls."
What Fugu doesn't try to do is replace those upstream models. If Anthropic's top systems are off the menu in Asia, the answer, in Sakana's architecture, is to route around them through everything else that's still accessible. That is a hedge against U.S. policy, even if Sakana's spokespeople will not call it that.
Qihoo 360's Tulongfeng, unveiled on June 24, takes a more direct approach. It is positioned as a cybersecurity AI tool comparable to Mythos, and Insurance Journal reported on the launch alongside the Sakana release as part of the same pattern. Tulongfeng's capability claims are still company-sourced; independent benchmarks are not yet public. The right read on Tulongfeng is that 360 is bidding for the procurement dollars Mythos can no longer capture in China, not that it has matched Mythos head-to-head.
The thing worth watching is whether orchestration products like Fugu actually sell, or whether procurement gravity simply shifts to whichever single vendor is willing to take the export risk. Sakana's pitch depends on buyers accepting a more complicated stack: a router, plus a basket of upstream models, in exchange for not getting locked out again. That is a real ask, and Ito has been candid that "U.S. models remain important to Asia." That line is the operative constraint: the workaround is a workaround, not a replacement.