Japan's Fugu Ultra AI routes around export controls and rivals top AI models

Tokyo startup Sakana AI released Fugu on June 22, bundling multiple AI models into a single API that the company says matches the performance of Anthropic’s Fable 5 and Mythos Preview.

“Our ‘Fugu Ultra’ model matches the performance of Fable and Mythos, delivering frontier capability without the risk of export controls,” wrote Sakana AI in a post on X.

According to the release, Fugu is a trained orchestrator that determines when to respond to a query independently and when to delegate components of a job to specialized models from external sources. It internally oversees model selection, task allocation, and response synthesis.

Sakana Fugu has two tiers:

• Fugu.
• And Fugu Ultra.

Fugu is built for everyday use, like coding, chat, and code review workloads. The model prioritizes lower latency. Fugu Ultra handles more complex, multi-step problems by coordinating a larger pool of expert agents. Users can gain access to Fugu models via one OpenAI-compatible API.

Introducing Sakana Fugu: A full multi-agent orchestration system accessible via a single model API.

Our ‘Fugu Ultra’ model matches the performance of Fable and Mythos, delivering frontier capability without the risk of export controls.

Try it: https://t.co/aDEFyySWlS 🐡 pic.twitter.com/43wzMAhyzT

— Sakana AI (@SakanaAILabs) June 22, 2026

Fugu Ultra positions itself as an alternative to export-restricted models

Sakana made benchmark tables that show how Fugu Ultra, Fable 5, and Mythos Preview stack up in terms of code, thinking, science, and agentic tasks. The company made it clear that neither Fable 5 nor Mythos Preview is in Fugu’s agent pool because the public is not allowed to see them.

Some countries can’t access top American AI models due to export controls, which is one reason Sakana went ahead with the launch of Fugu.

“Fugu stands shoulder-to-shoulder with leading models like Fable and Mythos across the industry’s most rigorous engineering, scientific, and reasoning benchmarks,” the company wrote on X. The post has over 29,000 likes and ~7,600 reposts within hours.

Tokyo was pushing to reduce dependence on foreign AI providers. Japan’s Digital Minister Hisashi Matsumoto warned earlier this month that the country risked becoming an “AI colony” without faster domestic development. Within the ruling Liberal Democratic Party, secretary general Akihisa Shiozaki argued in May that Japan should prioritize diversifying its AI suppliers over building sovereign systems from scratch.

Sakana did not train a single huge foundation AI model to directly compete with American companies. The Japanese startup created a route layer that sees different sources as replaceable parts. If one AI service limits access, the system will switch to another provider.

Sakana AI has published peer-reviewed work on learned model orchestration, including two papers titled Trinity and Conductor. The research papers support Fugu’s approach.

The company ran a beta program with about 500 users before the public launch. One software engineer quoted in the release said that Fugu Ultra “gives comprehensive answers and finds the bugs others miss” during code review, surfacing more than 20 issues where competing tools flagged roughly three.

Japan’s government has backed domestic AI efforts with subsidies and infrastructure funding. SoftBank, Sakura Internet, and local chipmakers have received government support to build computing capacity and local models. Tokyo also joined Washington’s Genesis Mission with a $1 billion, five-year AI investment commitment announced in June.

Sakana clarified that in their performance comparisons, metrics for non-Fugu models are based strictly on self-reported data from the respective creators and not validated by neutral, third-party entities.

If you're reading this, you’re already ahead. Stay there with our newsletter.