When Washington forced Anthropic to switch off its most powerful models worldwide on June 12, the move was meant to deny adversaries access to frontier capability. Within ten days, two companies outside the United States had offered a pointed reply: ban the best model, and they would rebuild its performance out of the parts left lying around.
The first answer came from Tokyo. On June 22 the startup Sakana launched Fugu, named for the pufferfish, a system it says matches the restricted Claude Fable 5 and Mythos on benchmarks without touching either. Fugu is itself a model trained to call other models, pooling Google's Gemini 3.1 Pro, OpenAI's GPT-5.5 and Claude Opus 4.8 and combining them for each task. On Sakana's own tests, the flagship Fugu Ultra posted 73.7 on SWE-Bench Pro, ahead of Opus 4.8's 69.2, and 95.5 on the GPQA-D science exam. CEO David Ha framed it less as a trick than a doctrine: "Human intelligence is fundamentally a collective intelligence." Relying on a single vendor for national infrastructure, he argued, is a massive risk when access can vanish overnight. "By orchestrating the world's models, we are delivering the resilient blueprint required for AI sovereignty."
The catch is built into the idea. An orchestrator can only be as good as the models it can reach, and Fugu still trails the banned Fable 5 on the hardest tasks. Prime Intellect engineer Elie Bakouch dismissed it as a closed orchestrator built on closed models, leaving users in control of even less than before. The pricing is not gentle either: Fugu Ultra runs about $35 per million tokens, level with GPT-5.5, and the service is live everywhere except the EU, where Sakana is still squaring its routing with GDPR.
Beijing's answer was more direct: build a frontier model and give it away. Zhipu, which trades as Z.ai, released GLM-5.2, a 744-billion-parameter open model with no usage restrictions. It ranked second in the world behind Fable 5 on CodeArena's front-end coding board and pushed Zhipu to third place on the Artificial Analysis Intelligence Index, behind only Anthropic and OpenAI, the first time a Chinese lab has cracked the global top three. Jefferies called it "a milestone for Chinese AI," and Zhipu's Hong Kong shares jumped as much as 42 percent intraday. When Musk predicted on X that a Chinese rival to Fable 5 was probably a quarter away, CEO Tang Jie replied simply, "won't take that long," adding that cutting-edge intelligence "should be open, available, extensible," and never "withdrawn at any time."
The two responses share an argument and split on method. Sakana says you do not need the banned model if you can conduct the ones still standing; Zhipu says you do not need it because an open one is already closing in, for free. Both are bets that capability cannot be cleanly fenced. Restrict the frontier, and the market routes around the restriction, through orchestration in Tokyo and open weights in Beijing. The export order assumed the best model was a chokepoint a government could hold. Sakana and Zhipu are wagering it is already closer to a commodity, and that the harder Washington squeezes, the more of the world it gives a reason to prove the point.