On June 12, the US government ordered Anthropic to immediately disable access to two of its most powerful AI models for all foreign nationals, including Anthropic's own non-American staff. Because verifying nationality per customer in real time is not practical, Anthropic shut the models off for everyone globally. Businesses that had built internal tools or workflows on those models lost access the same day, with no timeline for restoration.
Ten days later, Sakana AI, a Tokyo company valued at $2.65 billion, launched a product called Fugu. The pitch was direct: a single AI service that routes your request across a pool of different AI models, picks the best combination for your task, and returns one answer. If any provider restricts access or goes offline, the pool adjusts.
Sakana was founded in 2023 by David Ha, formerly head of research at Stability AI, and Llion Jones, one of the original co-authors of the academic paper that made modern AI possible. The company has raised around $379 million in total, with backers including Nvidia, Mitsubishi UFJ, Khosla Ventures, and In-Q-Tel, the venture arm linked to US intelligence agencies. It is Japan's most valuable private technology startup.
Fugu comes in two versions. The standard version handles everyday tasks: chat, code review, quick analysis. Fugu Ultra targets harder work: multi-step research, security assessment, patent investigation. Both use the same single connection point, which means businesses can plug Fugu into existing tools without rebuilding anything.
On pricing, Fugu Ultra starts at $5 per million words of input and $30 per million words of output. For very long tasks above a certain length, those rates roughly double. There is a subscription option starting at $20 a month for lighter use. Sakana also charges a single blended rate rather than stacking fees for every model that runs inside a given request, which makes costs more predictable.
The benchmark claims Sakana published are impressive on paper, but they come from Sakana itself and have not been independently verified. In real-world tests by outside users within the first 24 hours of launch, the picture was more complicated. Some tasks reportedly took 30 minutes to complete. One comparison test found Fugu Ultra finished a software build project in 22 minutes at a cost of around $7, while a single leading model took 79 minutes and cost nearly $38. Faster and cheaper, yes, but the tester still judged the single model's output to be better quality.
The deeper criticism is structural. One research engineer put it plainly: Fugu is a closed system running on top of other closed systems. Before, you did not control the model. Now you do not even know which model handled your request or how. Sakana offers enterprise users the ability to exclude specific providers from the routing pool for compliance reasons, and lets them opt out of having their prompts used for future training. But the routing logic itself remains invisible.
For businesses in Europe, the service is not available at all yet. Sakana is still working through data compliance requirements under EU privacy rules, with no announced date for when that will change.
The broader story here is not just about one product launch. The US government's decision to pull Anthropic's models without warning demonstrated something that many businesses have not yet fully absorbed: AI access can be treated as a geopolitical lever, not just a commercial service. Any organization that depends on a single AI provider, especially one based in the US, now carries a form of supply chain risk that did not exist two years ago.
Fugu does address that risk, genuinely. If Anthropic goes offline again, or if another provider faces restrictions, a routing service can adjust. The catch is that you are trusting Sakana to make those adjustments correctly, on your behalf, without showing you the work. Whether that trade is worth making depends on how much you value availability over transparency.