Fugu, Sakana AI's Japanese bet

⏱️ Reading time: ~5 minutes

Introduction

On June 22, 2026, the Japanese startup Sakana AI announced Fugu, a new artificial intelligence model that promises to rival the most advanced systems on the market. According to the company, Fugu delivered results equivalent to those of Anthropic's Fable 5 and Mythos 5 models on the industry's most rigorous benchmarks, and is capable of handling complex multi-step tasks. More than a simple competitor, Fugu draws attention for proposing a different way of building AI — and for arriving at a moment when the geopolitical landscape of technology is tenser than ever.

Content

Fugu's big differentiator is not being a language model trained from scratch with trillions of parameters, as the American and Chinese giants do. It is, in practice, a multi-agent orchestration system trained to work with several different language models, bringing together various technologies that act as a collaborative intelligence. Instead of relying on a single giant model, Fugu works like a "conductor": when it receives a task, it can solve it directly or gather, select, and delegate parts of the work to specific models, according to the complexity involved.

This approach has both a technical and a strategic justification. Sakana AI argues that the method improves results without the need to train enormous models, reducing costs — an alternative to the race for more parameters and more billion-dollar investment, which today is concentrated in the hands of a few companies. The startup itself sums up the philosophy behind the product as a way to sidestep dependence on a single technology supplier, since the system orchestrates multiple agents that can be swapped for one another.

There is also a relevant geopolitical component. In the same announcement, the company highlighted that Fugu offers cutting-edge capabilities without the risk of export controls, a direct reference to the block imposed by the United States on Anthropic's technologies, an episode that led to the temporary suspension of access to models like Fable 5 and Mythos 5. In other words: part of Fugu's commercial appeal is precisely being an alternative "free" of that kind of international regulatory restriction.

In terms of performance, Sakana AI claims the model posted results comparable to Anthropic's on the SWE-Pro (software engineering), GPQA-D (complex scientific reasoning), and ALE-Bench (advanced problem solving) tests, plus good performance in automated data-science research with little human intervention. Commercially, Fugu is offered via an OpenAI-compatible API and comes in two versions: the standard one, aimed at day-to-day tasks with a balance between performance and low latency, and Fugu Ultra, focused on maximum answer quality for complex, multi-step problems.

Conclusion

Fugu represents a different bet within the global AI race: instead of trying to win on model size, Sakana AI bets on intelligent coordination between systems that already exist. It's a strategy that cuts costs and works around regulatory barriers, but it also raises a legitimate question — Fugu is not, technically, a language model of its own but an orchestrator of third-party models. That doesn't diminish its practical usefulness for companies, but it puts into perspective how much it represents (or doesn't) an independent technological advance by Japan against powers like the United States and China.

To learn more about the topic, see the blog: https://nascimentoab.com.br/blog/fable-5-mythos-5-two-brothers-ch-2-the-well-behaved-one