Two weeks after the US Commerce Department cut off global access to Anthropic’s Mythos and Fable 5 models, Tokyo’s Sakana AI launched Fugu, a multi-agent system it markets as delivering “frontier capability without the risk of export controls.” China’s 360 unveiled a rival tool, Tulongfeng, that it claims matches Mythos. Anthropic’s run-rate revenue topped $47 billion in May 2026.
Both launches reach for the same gap: governments and firms wary of depending on US AI they could lose overnight. Sakana co-founder Ren Ito still calls the situation a moment, not a permanent break.
An export control does two things at once. It denies a capability to an adversary, and it tells everyone else they are one policy memo away from losing the same tool. The second effect is the one reshaping Asia’s AI market right now.
On June 12, 2026, the US ordered Anthropic to switch off Mythos and Fable 5 for all foreign users. Within two weeks, Tokyo-based Sakana AI shipped Fugu, a model it describes as comparable to Fable 5, and China’s 360 Security Technology announced Tulongfeng, which it claims rivals Mythos. The timing reads like opportunism. The deeper signal is that frontier AI is fragmenting along the lines Washington just drew.
What changes when a model can be revoked is not the technology. It is who gets to decide whether you keep it.
The chokepoint moved from the lab to the licence desk
Fugu is not a bigger chatbot. It is what Sakana CEO David Ha calls an “orchestration model” — a system that coordinates several specialized models through APIs rather than trying to do everything itself. “Orchestration Models are the next frontier, beyond bigger models,” Ha said. Sakana presented the underlying research, on a technique it calls evolutionary model merging, at the ICLR 2026 conference.
That architecture matters because of the threat it is built against. “Access to top models can disappear overnight,” Ha argued, calling collective intelligence “the practical hedge against this concentration of power.” A system that routes across many providers is harder to switch off than one that depends on a single American lab.
For Western readers, Fugu plays a role similar to Anthropic’s Claude-based agent stacks, but tuned for Japanese language, local tools, and multi-agent workflows. 360’s Tulongfeng and its companion tool Yitianzhen, by contrast, resemble specialized red-teaming and security-automation software built on frontier models. The crucial difference is jurisdiction: these systems are architected to avoid US export leverage entirely.
Here is the line most coverage will miss. If Fugu or Chinese security stacks become the default in local developer communities over the next twelve to eighteen months, future tools, plugins, and agent frameworks will be tuned to their APIs first. That makes the lock-in structural, not just commercial. The open question is whether the ban that triggered all this is temporary.
Three rulebooks, no shared standard
The fragmentation runs deeper than one ban. Each major market now governs frontier AI by a different logic, and the gaps between them are where regional providers grow.
China requires algorithm filing and security reviews under its Interim Administrative Measures for Generative AI Services, in force since August 2023, but stops short of the EU AI Act‘s detailed high-risk categories — leaving room for state-backed deployment of powerful models in security work. Japan leans on soft-law principles and benchmark evaluations from its AI Safety Institute rather than hard constraints. The US relies on sectoral rules plus the export controls now biting. Three rulebooks, no common floor.
This is why a model can disappear overnight and a market cannot snap back. Ito frames the moment as recoverable, and he may be right that US models stay important to Asia. But once governments build domestic AI as a hedge against revocation, the hedge does not unwind when the rule softens. The ban told Asia it was a guest in someone else’s system. Asia is now building its own house.
Beyond the headline
The power behind it
Control over frontier AI access is shifting from technical capability to export-licensing discretion inside a handful of governments. Compliance officers and trade lawyers, not just model researchers, now decide which systems can underpin national infrastructure. Asian startups are designing around that chokepoint, transferring leverage from Silicon Valley roadmaps to security doctrines in Washington, Beijing, and Tokyo.
The timing
This wave of launches lands as governments turn broad AI-governance rhetoric into concrete controls on specific models. Bans that once seemed hypothetical now hit production systems, forcing CIOs to revisit multi-year AI strategies mid-cycle. That policy-driven disruption creates a rare window in which lesser-known regional providers can win long-term platform positions.
The reach
For Western cloud and software vendors, the real damage is not lost direct sales but the secondary ecosystems built on non-US models. If Japanese or Chinese stacks become the local default, future plugins and agent frameworks will target their APIs first. Reclaiming mindshare gets harder even if export rules later soften.
What a revocable model means for your AI strategy
With Commerce expected to clarify AI export licensing in the coming weeks, anyone with East Asian exposure or a frontier-model dependency faces decisions now.
- Western enterprise CIOs
Map which of your cloud-hosted models could fall under licensing or country-specific restrictions. Review the US Bureau of Industry and Security materials at bis.doc.gov, and assume any single-provider dependency for critical systems is now a continuity risk, not just a vendor choice.
- Firms operating in Japan or China
If you sell into or run subsidiaries in these markets, examine Japan’s AI Safety Institute publications and China’s generative-AI measures before committing to a regional provider. Algorithm filings and security reviews can gate deployment by months.
- Technology investors
Watch whether Commerce includes carve-outs for allies in its next notice. Carve-outs signal Washington wants coordination over decoupling; silence signals Asian providers will double down on “export-control-free” branding and regional subsidy schemes will accelerate.
Explainer
- Orchestration model
- An AI system that coordinates several specialized models through APIs rather than relying on one large model for every task. The approach trades raw single-model scale for robustness, since work can route across multiple providers. Sakana AI’s Fugu is built on this design specifically so that losing access to any one model does not collapse the whole system.
- 360 Security Technology
- A Chinese cybersecurity firm founded by Zhou Hongyi, best known domestically for antivirus and internet-security products. It has moved into security-focused AI, including the Tulongfeng and Yitianzhen tools announced in June 2026. Zhou describes AI-driven vulnerability discovery as a national strategic asset, warning that asymmetric access could leave countries exposed in cyber conflict.
- EU AI Act
- The European Union’s risk-based law for artificial intelligence, politically agreed in 2023 and entering phased application from 2025. It created a specific category for high-impact general-purpose models, imposing transparency and risk-management duties on frontier providers serving Europe. Because compliance is required for market access, its obligations could become the default global standard for any lab with European users.
- AI Safety Institute
- Japan’s evaluation body, launched under the Digital Agency in 2024 to assess large language models on safety and reliability. It began publishing benchmark evaluations in 2025, giving domestic models a reference point without a binding AI statute. Its soft-law approach contrasts sharply with both the EU’s hard rules and Washington’s export controls.
RELATED STORIES
