Sakana AI has released a new system called Sakana Fugu that acts as a smart dispatcher for AI tasks. Instead of using one all-purpose AI model, Fugu routes different types of problems to specialized AI systems — some better at writing code, others better at math or science.

Think of it like a medical clinic with different specialists: when a patient comes in, the receptionist doesn't send everyone to the same doctor. A person with a broken arm goes to the orthopedist; someone with a skin condition goes to the dermatologist. Fugu does the same thing, but with AI models. It sends coding problems to AI trained heavily on code, math problems to AI trained on formal reasoning, and so on.

The system launched in beta in April 2026, and became widely available in June 2026.

What Fugu Does

Fugu coordinates multiple specialized AI systems and decides which one should handle each incoming task. Teams access it through an API that works the same way as OpenAI's system, which means companies already using OpenAI's tools can switch to Fugu without rewriting their code.

Sakana claims Fugu performs better than a single AI model on coding, math, and scientific problems — the exact types of tasks where using specialists has historically worked best. The strategy makes sense on paper: route different problem types to different strengths.

The real test is whether Fugu stays fast. When a task bounces between multiple AI systems, each handoff adds delay. Whether Fugu's routing logic is smart enough to keep overall response times competitive with using a single fast model is something real-world use will determine.

Why This Matters

No single AI model is best at everything. A system smart enough to use different models for different jobs should outperform any single model. That idea is well-established in AI research, but Sakana is selling it as a finished product you can pay for and rely on. Most other orchestration systems leave developers to build this routing logic themselves.

By using the same API format as OpenAI, Fugu makes it easier for teams to start using it — they don't have to rebuild monitoring systems or change how they manage AI prompts.

Teams thinking about using Fugu for serious work should ask some questions. The jump from beta to final release happened in just two months. For work that can't afford failures — like automated code review or scientific analysis — teams should understand what improved during that time and what happens when something goes wrong.

Sakana AI built its reputation on research into how multiple AI models can work together. Fugu is their first commercial product in that space, and it arrives as companies are investing heavily in multi-agent AI systems. For teams already using multiple AI models, Fugu offers a managed solution instead of building custom connectors. The trade-off is convenience against relying on one company's routing system — a decision that varies depending on team size and risk tolerance.