Sakana AI’s Fugu Takes the Fight to Global Tech Giants

Tokyo-based Sakana AI is challenging the industry’s obsession with monolithic models by launching a multi-agent system that orchestrates existing tech to match frontier performance.

The race for artificial intelligence supremacy has long been a game of brute force, defined by who can build the largest, most expensive foundation model. Tokyo-based Sakana AI is now betting that the future isn't about building a bigger monolith, but about building a better conductor. With the launch of Sakana Fugu, the startup has introduced a system that treats multiple high-performance models as an "agent pool," routing complex queries to specialized AI to get the best possible result.

The Sakana Fugu service, available in both standard and Ultra versions, functions as a master general contractor. When a user sends a prompt, the Sakana system decides which combination of AI tools—including versions of itself—can best solve the problem, handling everything from planning to final execution. This orchestration happens behind the scenes, allowing the user to interact with the system via a single, OpenAI-compatible API, effectively hiding the complexity of a multi-agent workflow.

Benchmarks and Performance

The performance figures are turning heads in the Asia tech scene. On the demanding SWE-Bench Pro, the Fugu Ultra model achieved a score of 73.7, comfortably outperforming industry standards like GPT-5.5 and Claude Opus 4.8. In specialized tests like LiveCodeBench, the system logged a 93.2, and it even managed to solve a Rubik’s Cube in 19 steps—the fewest among a field of top-tier competitors. While Sakana notes that these benchmarks exclude Anthropic’s Fable 5 and Mythos, the startup maintains that its system sits shoulder-to-shoulder with these frontier models in real-world application.

Why it Matters

This shift toward orchestration is a strategic pivot for the Japan-based startup, which is currently valued at over $2.5 billion. By moving away from single-vendor dependency, Sakana is addressing a growing geopolitical and operational anxiety: what happens to a business or a nation if it relies entirely on one foreign provider's API? Sakana AI suggests that a collaborative ecosystem is not just more efficient, but more resilient. If a specific model in their pool fails or becomes unavailable, the system can simply route the task elsewhere.

The Bigger Picture

The emergence of Sakana Fugu highlights a maturing AI market that is moving past the initial hype of massive, general-purpose engines. We are entering an era of "intelligent routing," where the value lies in knowing which tool to use for a specific job rather than just having the biggest one. For enterprises, this offers a middle path: the ability to access frontier-level performance without the prohibitive costs or the risks of locking their entire infrastructure into one proprietary platform. As the industry grapples with power constraints and the limits of scaling, this "team-of-models" approach may well become the new standard for efficiency.