Sakana AI's Fugu and Alibaba's Qwen fight the same assumption from opposite directions: US labs still lead, but that lead is no longer protected by size alone.
The AI race no longer looks like a straight contest over who can train the biggest model. In the same week, a Tokyo startup made a serious case for orchestration over scale, while Anthropic accused Alibaba of using 28.8 million Claude exchanges to pull US model capabilities into Qwen faster than normal research would allow.
If you build with these systems, that combination should get your attention. It says the next advantage may come from routing, specialization and access, not only from another giant training run in an American data center.
On June 19, Sakana AI published its Fugu technical report on arXiv, with a revised version on June 23. The Tokyo company describes Fugu as a family of orchestrator models that doesn't try to answer every query alone. Instead, it decides which frontier agents to use, how to structure the work, and when to combine their outputs. Sakana released two versions: Fugu, built for lower latency, and Fugu-Ultra, built for harder problems where quality matters more than speed.
The benchmark claims are the sharp part. Sakana says Fugu-Ultra reaches state-of-the-art results against publicly accessible models across software engineering, reasoning and scientific tests, including SWE-Bench Pro, Terminal Bench, LiveCodeBench, GPQA-Diamond, Humanity's Last Exam and CharXiv Reasoning. The company's own report frames the system as "shoulder-to-shoulder" with Anthropic's Fable 5 and Mythos Preview, even though those models are not in Fugu's worker pool because they aren't publicly accessible.
That matters to you because Fugu is not another story about a lab throwing more GPUs at the same problem. Sakana is making a different bet: a smaller coordinator can squeeze more capability out of existing models by choosing the right worker for the right task. It is a foreman model, not a bigger laborer.
Sakana is not some anonymous research shop posting a leaderboard trick. The company was founded in 2023 by David Ha, Llion Jones and Ren Ito, and it has already drawn backing from Nvidia, Khosla Ventures and major Japanese financial groups. Its broader research has long leaned toward collective intelligence, including model merging and systems built from many smaller parts. Fugu fits that pattern cleanly.
Preferred Networks is making a quieter Japanese argument from another direction. Nikkei Asia has reported that its PLaMo model, trained natively on Japanese data, can offer performance comparable to OpenAI's models at less than half the price. In June, Preferred Networks also announced a business alliance with Mitsubishi Heavy Industries to develop AI for mission-critical infrastructure and national security applications.
These aren't chatbot stunts. They are Japan's attempt to build AI capacity where language, cost and industrial reliability matter more than Silicon Valley branding. If you're a company buying AI tools in Tokyo, Osaka or Nagoya, a model that handles Japanese natively and costs less is not a patriotic detail. It's the procurement case.
The Chinese side of the story is rougher. According to reporting from Bloomberg, the Financial Times and Business Insider, Anthropic wrote to the US Senate Banking Committee accusing Alibaba-affiliated operators and its Qwen lab of conducting the largest known distillation campaign against Claude to date. Anthropic said the operators used nearly 25,000 fraudulent accounts and generated 28.8 million Claude exchanges between April 22 and June 5, 2026.
Anthropic said the campaign targeted commercially valuable capabilities including software engineering, agentic reasoning and long-horizon tasks. Alibaba has not publicly answered the allegation. That absence belongs in the story, because this is still an accusation from one company, not a court finding.
Still, don't miss the signal beneath the fight. Distillation is not magic. It is the practice of training one model on the outputs of another model, and it can be legitimate when done with permission. Anthropic's accusation is that Alibaba used Claude at industrial scale, through fake accounts, to transfer capability into Qwen without paying the normal cost of developing it independently.
Frankly, that is not the behavior of a market that thinks US models are untouchable. It is the behavior of a market that thinks the gap can be closed, and is fighting over how quickly that happens.
Qwen already has momentum. Public benchmark trackers have shown Alibaba's Qwen models competing near the top of several hard evaluation sets, while DeepSeek and other Chinese labs keep pressure on token pricing and open-weight availability. You do not have to accept every leaderboard as gospel to see the pattern. Chinese labs are no longer asking whether they can enter the frontier conversation. They are arguing over the fastest route through it.
The real pressure on American labs is economic. If Sakana can use orchestration to get frontier-like results from a coordinated pool of models, and Preferred Networks can undercut Western pricing with language-native systems, and Chinese labs can close gaps through cheaper open-weight releases or alleged distillation, then the old moat looks thinner. Massive compute still matters. But it is not the whole moat anymore.
That is the point worth carrying away from this week. The AI pecking order is not being rewritten by one spectacular model launch. It is being worn down by practical alternatives: routing instead of brute force, local language training instead of translation, and aggressive capability transfer where the rules are still being fought over in Washington.
Also read: Washington's AI export controls are handing Asian labs the frontier race they were supposed to lose • 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
