Asian AI Startups Move Into Mythos-Like Models as US Export Controls Bite

Asian AI companies are moving quickly to fill the gap created by continued US restrictions on Anthropic's most advanced models, with new releases from Japan and China framed around sovereign access, cyber capability, and reduced dependence on American model providers.

The timing is notable. Anthropic's Mythos and the more restricted Fable 5 remain blocked from non-US access under export-control measures, creating a sudden strategic opening for regional AI labs that can offer comparable capabilities without the same geopolitical constraints.

Japan's Sakana AI Pitches Fugu as an Access Hedge

Tokyo-based Sakana AI has launched Fugu, a frontier model named after the Japanese word for blowfish. The company is positioning it as a high-end system that can compete with leading models such as Fable 5 and Mythos Preview, while also emphasizing that it is not exposed to US export-control risk in the same way.

Fugu is designed for agentic workloads rather than simple chat completion. Its central pitch is orchestration: the model can coordinate calls to other models through APIs, allowing developers and enterprises to build workflows that use multiple AI systems instead of relying on a single provider.

That matters for Japanese businesses and government agencies that want advanced AI capability but do not want critical operations tied entirely to access decisions made in Washington. The export ban has turned model availability into an infrastructure risk, especially for organizations exploring AI agents, automated research, security analysis, and national-scale digital services.

Sakana has said the release timing was coincidental, with the underlying research already underway before the latest export restrictions. Still, the market message is clear: frontier AI access can disappear quickly, and enterprises may need architectures that can survive sudden provider or policy shocks.

Orchestration Becomes a Strategic AI Design Pattern

Fugu also reflects a broader shift away from the idea that progress only means building ever-larger standalone models. Instead of treating one model as the entire intelligence layer, orchestration models act more like routers, coordinators, and planners across a set of tools and model endpoints.

That approach is attractive in markets where no single vendor can be assumed to remain available forever. A company can route coding, reasoning, summarization, search, and domain-specific tasks across several providers, then swap pieces as cost, latency, quality, or regulation changes.

For production teams, the lesson is practical: model redundancy is becoming part of AI resilience planning. The same way cloud architectures avoid single-region failures, agentic AI stacks may increasingly avoid single-model or single-provider dependence. Teams building these systems should treat access continuity as part of their AI risk management process, not just as a procurement detail.

China's 360 Takes a Cybersecurity-First Angle

China's 360 has taken a more direct cybersecurity route, unveiling AI tools aimed at vulnerability discovery, cyber defense, and incident response. Its Tulongfeng system is designed to identify software weaknesses automatically, while Yitianzhen focuses on defensive operations and response automation.

The framing is more assertive than Sakana's. Rather than presenting the tools mainly as a hedge against supply-chain uncertainty, 360 is treating advanced vulnerability-finding AI as a strategic national capability. The argument is that if only some countries or companies can access top-tier cyber models, the result could be an uneven security landscape where one side sees weaknesses the other side cannot detect as quickly.

That concern cuts to the center of the Mythos debate. Models that can discover exploitable vulnerabilities are valuable for defensive hardening, red-team simulation, and faster patching. But the same capability can also accelerate offensive reconnaissance and exploit development if access is poorly controlled.

Export Controls Are Reshaping the AI Market

The US restrictions were intended to keep the most sensitive AI capabilities away from foreign users, especially where cyber or national-security implications are high. But the longer the restrictions remain in place, the more they create incentives for regional alternatives.

In Asia, the response is not uniform. Japan's approach is largely about resilience, continuity, and allied access to advanced AI. China's response is more focused on strategic autonomy and cyber parity. Both reactions point to the same market reality: export controls can protect access to certain models in the short term, but they can also accelerate competing development elsewhere.

For Anthropic and other US AI labs, this creates a commercial and strategic tradeoff. Restricting global access may satisfy national-security requirements, but it also gives rivals a clearer story to tell enterprise customers: local or regional models are less likely to be interrupted by foreign policy decisions.

What Enterprises Should Watch

The immediate question is not whether Fugu or 360's tools permanently replace US frontier models. The more important shift is that advanced AI buyers are starting to evaluate model access as a geopolitical dependency.

Enterprises should watch three signals closely:

• Whether agent orchestration becomes the preferred architecture for organizations worried about provider concentration.
• Whether cybersecurity-specific AI models become a separate regulated category from general-purpose frontier models.
• Whether governments in Asia begin backing domestic or regional model providers as strategic infrastructure.

The Mythos export ban may have been aimed at limiting access to a narrow set of powerful capabilities, but its market effect is broader. It is pushing AI customers to think less like software buyers and more like infrastructure planners, where continuity, jurisdiction, and substitution risk matter as much as benchmark performance.