Generated by Codex with GPT-5

Techmeme surfaced this June 27, 2026 story in its cluster on Sakana Fugu and 360 Tulongfeng, and the concrete original URL is Kate Park’s TechCrunch article, Asian AI startups launch Mythos-like models as Anthropic’s export ban drags on. The important point is not that two Asian AI companies made ambitious benchmark claims. It is that U.S. model-access restrictions are already being converted into a commercial and geopolitical pitch for alternatives.

What Happened

TechCrunch reported that two companies moved into the space opened by the U.S. directive limiting access to Anthropic’s Fable 5 and Mythos 5. In China, cybersecurity company 360 unveiled Tulongfeng, a vulnerability-discovery model it says can compete with Mythos, plus Yitianzhen, a tool for cyber defense and incident response. In Japan, Sakana AI launched Fugu and Fugu Ultra, presenting them as frontier-level systems that can avoid dependence on any single U.S. model provider.

The timing matters because Anthropic said on June 12 that the U.S. government had ordered it to suspend access to Fable 5 and Mythos 5 for foreign nationals, including foreign national Anthropic employees, after concerns about cyber misuse. Anthropic disagreed with the action but complied, disabling access broadly so it could stay within the directive. OpenAI’s GPT-5.6 rollout then landed in the same political weather, with limited early access and public concern that a government access process could become the default for the strongest models.

Sakana’s Fugu launch takes that uncertainty and turns it into product positioning. Fugu is not described as a single monolithic model trying to beat every other model on every task. It is an orchestration model: a model trained to decide when to answer directly, when to delegate to other models, how to coordinate those agents, and how to synthesize their work behind one API. Fugu Ultra is aimed at harder, multi-step tasks such as research, code review, cybersecurity analysis, paper reproduction, and patent investigation.

That makes the launch more than a local competitor story. Sakana is selling resilience. If one provider becomes unavailable because of export controls, policy shifts, pricing, or commercial conflict, the orchestrator can route around the disruption by swapping the agent pool. That is a technical architecture, but it is also an argument about national and enterprise risk.

Why This Is More Than Benchmark Theater

The obvious caveat is that the claims need to be treated carefully. Sakana says Fugu Ultra stands near Anthropic’s Fable 5 and Mythos Preview on demanding engineering, science, and reasoning benchmarks, but its release notes also say the comparison uses provider-reported scores for other models and that Fable and Mythos are not in Fugu’s agent pool because they are not publicly accessible. 360’s claims also arrive through the lens of strategic messaging around cybersecurity and national capability.

Even with those caveats, the market signal is real. Customers do not need every benchmark claim to be independently settled before they change procurement behavior. If access to a frontier model can disappear overnight, buyers in Japan, China, Europe, India, or the Middle East have a strong reason to ask whether their most important workflows should depend on a single American API.

This is where “AI sovereignty” becomes more practical than ideological. The useful version is not that every country must own a complete frontier stack from chips to model weights. It is that governments and companies want options: multiple providers, local-language competence, data controls, swappable model pools, and a way to keep critical workflows running if policy changes somewhere else.

Sakana’s positioning fits that version. Its pitch is not simply “replace U.S. models with Japanese models.” It is closer to “abstract over the model layer so no single provider has veto power over your workflow.” That is a compelling architecture for companies that want frontier capability but cannot tolerate sudden access, compliance, or jurisdictional surprises.

The Cybersecurity Twist

The cyber angle sharpens the tradeoff. Mythos became important because vulnerability discovery and remediation are among the first domains where frontier models appear to create visible, high-value leverage. A model that can inspect codebases, reproduce issues, find subtle vulnerabilities, and generate useful fixes is valuable to defenders. It can also make governments nervous because the same capability can lower the skill threshold for attackers.

360 framed vulnerability-finding AI as a strategic asset, warning about a world in which some actors can use advanced vulnerability discovery while others cannot. That is the security dilemma in miniature. If only a few trusted U.S. organizations can use the strongest cyber models, defenders elsewhere will look for substitutes. Some substitutes may be less aligned with U.S. policy goals, less transparent, or more directly tied to national industrial strategy.

So the restriction may reduce one class of immediate risk while increasing another: it encourages non-U.S. labs and security companies to build and market independent cyber-capable systems as a matter of strategic necessity. That does not mean the U.S. should release every model broadly. It does mean model-control policy has second-order effects, and those effects arrive quickly.

What To Watch

The key question is whether this becomes a permanent architecture shift. If the Anthropic and OpenAI access disputes are treated as temporary crises, buyers may return to the strongest U.S. models as soon as access stabilizes. But if labs and governments keep handling frontier releases through improvised, model-by-model access negotiations, customers will increasingly design around uncertainty.

That would make orchestration, routing, and provider substitutability central product features. The winners would not only be the labs with the best single model. They would be the companies that can keep workflows reliable across a shifting pool of proprietary, open, regional, and domain-specialized models.

The durable takeaway is that access policy is now part of model capability. A model that scores higher but cannot be used consistently is less useful than a slightly weaker system that is available, compliant, and resilient. The Fugu and Tulongfeng launches show how fast the rest of the market will turn that lesson into products.