Sakana Fugu: Features, API, Performance, and Reviews of the Multi-Agent Orchestrator

‍Sakana Fugu is a multi-agent artificial intelligence system developed by the Japanese laboratory Sakana AI. Rather than relying on a single large language model, it selects and coordinates several specialized models to solve a task behind a single interface.

Publicly unveiled on June 22, 2026, Fugu specifically targets programming, complex reasoning, scientific research, and issues requiring multiple verification steps. It distinguishes itself from Claude, GPT, or Gemini by its primary function: it is not just an additional model, but an orchestration layer capable of utilizing different agents.¹

What is Sakana Fugu?

Sakana Fugu is a family of orchestrator models offered by Sakana AI, an artificial intelligence research laboratory based in Tokyo.

The word fugu also refers to the famous Japanese pufferfish. In the context of this article, however, it is not about the fish, but a multi-agent AI system marketed as an API.

Fugu receives a request like a traditional model. In the background, however, it can:

• analyze the type of problem posed;
• select one or more suitable models;
• assign them roles or sub-tasks;
• organize their interactions;
• verify and synthesize the results.

For the user, this orchestration remains largely invisible. They send a single request and receive a final response, without having to manage multiple providers, API keys, or agent chains themselves.

This approach is sometimes presented as a mixture-of-agents framework, by analogy with so-called mixture of experts. The difference is that Fugu coordinates complete models accessible as agents, instead of only activating internal sub-networks within a single model.

How does Sakana Fugu work?

Fugu itself is a language model trained to understand a request and build an agent organization tailored to that request.

It therefore does not rely solely on a list of manually written rules. The system learns to decide which model to engage, in what order, and with what instructions.

An adaptive orchestration process

In a simplified use case, the process can be summarized in five steps:

1. Fugu interprets the request and its level of difficulty.
2. It selects the agents likely to be the most competent.
3. It builds a suitable collaboration process.
4. The agents execute, critique, or complete the work.
5. Fugu aggregates the contributions into a final response.

For a simple question, this organization can remain lightweight. For a complex problem, Fugu Ultra can coordinate up to three agents and produce a longer process.

This adaptive task delegation It avoids systematically using the most expensive mechanism. The number of agents and the level of reasoning can vary depending on the detected difficulty.

Trinity and Conductor: The Technologies Behind Fugu

Sakana Fugu leverages two research papers presented at ICLR 2026: TRINITY and Conductor. The commercial product goes further than these prototypes, but they help understand the fundamental principles.²

TRINITY: Select a Model and Assign a Role

TRINITY uses a small coordinator model associated with a very lightweight selection head.

At each step, the coordinator selects a model and assigns it one of the following three roles:

The process can span multiple rounds. A first agent develops a plan, a second performs the calculations or writes the code, then a third verifies the solution.

TRINITY's coordinator has been optimized, in particular, with an evolutionary strategy called sep-CMA-ES. This method tests different coordination policies, evaluates their results, then retains and combines the most effective strategies.

It is thus a form ofevolutionary optimization applied not directly to the content of the responses, but to the way models are organized.

Conductor: Build a Natural Language Workflow

Conductor adopts a more open approach.

Instead of merely selecting a model and a role at each turn, it generates a genuine multi-agent workflow in natural language. It can define:

• the sub-tasks to be completed;
• the responsible agents;
• the instructions given to each;
• the structure of their interactions;
• the method for synthesizing the results.

Conductor is trained by reinforcement learning to discover effective collaboration strategies.

This architecture is an important part of Fugu Ultra, which must be able to handle longer workflows, use tools, and maintain a shared memory among multiple agents.

Fugu and Fugu Ultra: what are the differences?

At the time of its public launch, Sakana AI offers two models: Fugu and Fugu Ultra.

During the beta phase, the first version was called Fugu Mini. It was simply renamed Fugu upon commercialization.

Fugu is designed as the default model for interactive uses.

It is suitable for programming, code review, conversational assistants, and analyses where response time is crucial.

Fugu Ultra prioritizes depth and quality, at the cost of potentially higher latency and expense.

Sakana AI notably cites:

• the reproduction of scientific articles;
• data science competitions;
• cybersecurity analyses;
• documentary or patent research;
• complex problems involving multiple steps.

Fugu Ultra is therefore not necessary for every query. Its value becomes apparent primarily when the cost of a mediocre answer outweighs that of a longer reasoning process.

Is Sakana Fugu a model like Claude, GPT, or Gemini?

Fugu can be queried like a traditional model, but its nature is different.

Claude, GPT models, and Gemini are primarily foundation models developed by a provider. Fugu acts as an orchestrator model above a set of models and agents.

Fugu therefore doesn't necessarily replace Claude, GPT, or Gemini. It can use them as components within its own system.

The most accurate comparison isn't "Fugu versus Claude," but rather:

Should one select a model oneself, or entrust this selection and coordination to Fugu?

Is Sakana Fugu merely a model router?

A traditional router analyzes a query, selects a model, then transmits the task to it.

Fugu aims to go further. It can organize multiple interventions, assign different roles, and adapt the workflow based on intermediate responses.

In certain situations, the system can call upon one agent to develop a strategy, another to implement it, and a third to identify errors.

This multi-step reasoning synthesis represents Fugu's key differentiator.

Our review of Sakana Fugu

Sakana Fugu offers a particularly interesting approach to multi-agent orchestration.

Instead of requiring developers to manually select Claude, GPT, or Gemini and then build their own workflows, Fugu encapsulates this complexity behind a familiar API.

Its combination of dynamic routing, reinforcement learning, and evolutionary optimization is its main advantage. Its published results on code, scientific reasoning, and agentic tasks indicate that coordination can sometimes surpass the performance of each model used separately.

Nevertheless, several limitations prevent us from hastily concluding it's a revolution.

The routing remains proprietary, costs and latency are less predictable, independent evaluations are still rare, and the service is not available in the European Union at the time of its launch.

Therefore, Fugu seems particularly relevant for teams that already use multiple models and want to outsource their selection and coordination. An organization requiring full traceability or strict control over each provider might prefer to maintain its own orchestration layer.

Nevertheless, Sakana Fugu illustrates an important evolution: the improvement of AI performance no longer solely depends on ever-larger models. It can also come from better organization of existing intelligences.