Sakana Fugu: A Multi-Agent Orchestration System as a Foundation Model

（＊日本語は英文の後に）

We are excited to introduce Sakana Fugu, our flagship international commercial AI product—a multi-agent orchestration system, now opening applications for early beta testers. Sakana Fugu coordinates pools of frontier foundation models to achieve state-of-the-art performance across coding, mathematics, scientific reasoning, etc.

Initially, our Sakana Fugu model will be available as an API, where it has served as a key internal tool for our own researchers and engineers, and we are now ready to invite people outside Sakana AI to try it:

👉 Apply for Beta Test

Sakana Fugu Model, which is a small language model itself, learns to call LLMs (left). In the course of training, it can learn to call itself, enabling Test-time scaling (right). The actual coordination in Sakana Fugu is adaptive and complex. 1

Pushing the Boundaries by Collective Intelligence

A core conviction at Sakana AI is that the most capable AI systems will not be monolithic models scaled in isolation, but collections of specialized agents working together. This thread runs through everything we have built: evolutionary model merging, which showed that diverse open-source models can be combined to produce capabilities none possessed individually; The AI Scientist, which demonstrated that coordinated AI agents can autonomously execute the full cycle of scientific research; ShinkaEvolve, which uses evolutionary search over a pool of LLM-generated programs to discover algorithms that outperform human-written solutions; and AB-MCTS, which showed that multiple frontier models cooperating through tree search can substantially outperform any individual model on hard reasoning tasks.

Sakana Fugu is the product form of this research direction.

Sakana Fugu 🐡

Conventional approaches to utilizing foundation models often require users to manage multiple API keys, as models from different providers tend to specialize in distinct areas. This multi-model management leads to economic inefficiency. Moreover, since model strengths are frequently problem-specific rather than broad area-specific, fine-grained optimization through model switching is difficult for end-users.

Sakana AI’s Fugu models resolve these limitations. Fugu models achieve superior performance by dynamically coordinating and orchestrating a diverse pool of powerful models. Instead of using domain knowledge to prescribe team organization, roles, or workflows, Fugu learns to dynamically assemble agents from a pool and coordinate them through non-obvious but highly efficient collaboration patterns.

Sakana Fugu models are based on our ICLR 2026 papers ( Trinity and Conductor), and we have substantially further improved the methods to increase the performance and user experience, to be offered as a commercial product.

This adaptive, dynamic orchestration grants Fugu models superior performance on established benchmarks. The above table is a subset of our current results for our models in beta.

Using Sakana Fugu

Sakana Fugu is accessible via APIs, with compatibility for standard OpenAI-format endpoints. If you are already using GPT, Gemini, or Claude via API, Sakana Fugu can be integrated into existing workflows with minimal changes. Behind that familiar interface, Sakana Fugu handles coordination across the model pool automatically — establishing the collaboration topology, assigning the roles and dispatching the subtasks to complete complex tasks.

Two variants are available: Sakana Fugu Mini 🐟, optimized with latency in mind, and Sakana Fugu Ultra 🐡, the full orchestration system, optimized for performance for demanding tasks.

Join the Beta

We are looking for researchers and engineers from all areas to join as early testers. We want to understand how Sakana Fugu performs across domains we have not yet tested internally, where it falls short, and what researchers and engineers most need from a system like this.

If you are using foundation model APIs in coding assistants like OpenCode and Codex, or in your engineering, business-specific projects where you would like to see if Fugu models bring performance or novelty advantages, we would love to have you involved.

👉 Apply to Join the Beta

Publications

Xu, Sun, Schwendeman, Nielsen, Cetin, Tang. TRINITY: An Evolved LLM Coordinator. ICLR 2026.

https://arxiv.org/abs/2512.04695

Nielsen, Cetin, Schwendeman, Sun, Xu, Tang. Learning to Orchestrate Agents in Natural Language with the Conductor. ICLR 2026.

https://arxiv.org/abs/2512.04388

Footnotes

1. One result from our papers points to an interesting future direction. When a Fugu model is allowed to call itself recursively, reading its own prior output as context and deciding whether to revise its coordination strategy, a new form of test-time scaling emerges. The model recognizes when its first attempt fell short and spins up a corrective workflow. The depth of recursion becomes a tunable compute axis at inference time, requiring no retraining. A small model, by reading itself, can iterate toward answers that neither it nor any of its workers could reach in a single pass. ↩

2. Self-reported score with custom Anthropic scaffold. SWEPro were evaluated with the mini-swe-agent scaffold. However, we use the scores reported by Anthropic for Opus with the max thinking efforts due to frequent timeouts during our evaluation trials. ↩

3. 自己呼び出しがもたらす、新しい推論時スケーリング　Sakana Fuguが自分自身の出力を入力として読み込み、協調のしかたを見直しながら再帰的に自分を呼び出せるようにしたところ、新しいタイプの推論時スケーリングが現れることがわかりました。モデル自身が「最初の答えでは不十分だった」と気づき、修正のためのワークフローを自ら立ち上げるのです。再帰の深さは推論時に調整でき、再学習は必要ありません。小さなモデルであっても、自分自身の出力を読み返すことによって、1回の推論では到達できなかった答えへとたどり着けるようになります。 ↩