Sakana Fugu – A Multi-Agent Orchestration System by Sakana AI

1. What is Sakana Fugu

Sakana Fugu is a multi-agent orchestration system launched by Sakana AI that dynamically schedules top-tier models through a single API, enabling emergent collective intelligence. The system features a built-in TRINITY evolution-driven orchestrator and a Conductor reinforcement learning orchestrator, automatically assigning roles such as thinker, executor, and verifier to different models without requiring manual workflow presets. Fugu supports recursive self-invocation and flexible model pool management, offering both Standard and Ultra editions. It achieves cutting-edge performance on rigorous benchmarks, rivaling Fable 5 and Mythos Preview, while eliminating single-vendor dependency and export control risks.

Image source: Official article

Technical Positioning and Domain: Sakana Fugu belongs to the field of Multi-Agent Systems (MAS) and LLM orchestration, operating above the model routing layer. It is not a single model in the traditional sense, but rather a specially trained language model that understands when to delegate, how to communicate, and how to integrate, serving as an intelligent coordination layer between multiple models. The core innovation of this tool lies in elevating multi-model collaboration from hard-coded pipelines to a self-learning, dynamically evolving system, suitable for enterprise-grade applications requiring complex reasoning, multi-step collaboration, and high-reliability output.

Research Background: Sakana AI is a research institution focused on efficient AI models and agent systems, with its research rooted in theories of natural evolution and collective intelligence. The technical foundation of the Fugu system stems from two ICLR 2026 papers—"TRINITY: Evolution-driven Orchestrator for Multi-LLM Systems" and "Conductor: Reinforcement Learning for LLM Orchestration." The research motivation lies in addressing three common pain points in current multi-model systems: low efficiency of fixed pipelines, risks of single-vendor lock-in, and lack of dynamic adaptability. Sakana AI aims to combine evolutionary algorithms with reinforcement learning to enable model clusters to spontaneously emerge collaborative capabilities that surpass any single member.

Core Value: Sakana Fugu resolves the core contradiction in enterprise-grade AI applications—how to maintain the convenience of a single API call while gaining the performance improvements and fault tolerance brought by multi-model collaboration. Its value is reflected in three dimensions: first, performance breakthroughs, surpassing single frontier models on rigorous benchmarks such as MATH, HumanEval, and CTI-REALM; second, architectural flexibility, with a replaceable underlying model pool that automatically routes around unavailable single vendors, fundamentally avoiding service interruptions and export control risks; and finally, cost optimization, where the Fugu Ultra edition achieves equivalent or even superior reasoning performance at one-third the price of Opus 4.8.

Technical Features: The core technical feature of Sakana Fugu lies in its model-as-orchestrator architecture—the system itself is a lightweight language model that uses lightweight heads to predict the next model to invoke and its role from hidden states, achieving efficient routing with minimal additional computational cost. Its dynamic self-learning capability allows the system to adaptively adjust collaboration strategies based on task type (coding, math, reasoning, etc.) without manual presets. The recursive self-invocation mechanism enables Fugu to call its own instances as higher-level coordinators when handling complex tasks, forming a hierarchical reasoning architecture.

2. Key Features

• Unified API Orchestration: Users call Fugu via an OpenAI-compatible endpoint, and the system automatically handles model selection, task delegation, verification, and answer synthesis. This means developers do not need to install any proprietary SDKs or modify existing code architecture—simply changing the endpoint is sufficient for integration, significantly reducing integration costs. This design allows Fugu to seamlessly embed into various toolchains already adapted to the OpenAI API.

• Dynamic Role Assignment: Based on the TRINITY evolutionary coordinator, the system autonomously learns to assign roles of Thinker, Executor, and Verifier to different models. The Thinker is responsible for deep reasoning and strategic planning, the Executor focuses on specific task execution, and the Verifier performs quality checks on outputs. Unlike traditional fixed pipelines, this assignment is dynamic, with the system adjusting role allocation based on task type and real-time model performance to avoid efficiency bottlenecks.

• Recursive Self-Invocation: Fugu supports invoking its own instance as a coordinator to handle complex tasks requiring multi-round deep collaboration. For example, when completing a comprehensive research report, Fugu can recursively call itself to separately handle subtasks such as literature review, data analysis, and conclusion generation, with each subtask internally invoking expert models for collaboration. This hierarchical architecture enables the system to tackle complex problems that traditional single-call approaches cannot solve.

• Flexible Model Pool Management: Users can independently select or exclude specific vendor models in the console to meet data privacy and compliance requirements. This feature is particularly important for regulated industries such as finance and healthcare—enterprises can restrict sensitive data to local or compliant vendor models for processing while using other models for non-sensitive tasks, achieving a balance between security and performance.

• Dual Versions for Scenario Coverage: Fugu Standard balances performance and latency, suitable for low-latency scenarios like daily coding and conversation; Fugu Ultra coordinates a deeper expert pool, maximizing answer quality for high-difficulty tasks by introducing more expert models and more complex collaboration strategies. The Ultra version performs exceptionally well on benchmarks requiring deep reasoning, such as MATH and GPQA, but with correspondingly increased inference latency.

• Resilient Anti-Risk Architecture: The underlying model pool is replaceable, and when a single model is unavailable, routing automatically bypasses it to avoid service interruptions and export control risks. This architecture design gives Fugu inherent fault tolerance—even if a vendor's model becomes unavailable due to policy, service outage, or performance fluctuations, the system can maintain normal service through routing strategies, ensuring business continuity.

3. How to Use

1. Environment Requirements and Prerequisites: Using Sakana Fugu does not require local deployment; the system is provided as an API service. Users need a valid Sakana AI API key and must ensure network access to Sakana AI's service endpoints. Currently, Fugu supports all clients and frameworks compatible with the OpenAI API standard, with no additional dependencies required.

2. API Integration Configuration: In your existing workflow, change the API endpoint from OpenAI's https://api.openai.com/v1 to Sakana Fugu's endpoint address (link to be updated after official release). Also, replace the API key with the one provided by Sakana AI. This step only requires modifying a few lines in the configuration file, with no need to rewrite any business logic code.

3. Version Selection and Parameter Settings: Choose the version based on task complexity. Use Fugu Standard for low-latency scenarios like daily coding and conversations; select Fugu Ultra for high-difficulty research, complex mathematical reasoning, and similar tasks. Specify the version in the API request using the model parameter, e.g., model=fugu-standard or model=fugu-ultra. Additionally, standard parameters like max_tokens and temperature can be used to control generation behavior.

4. Sending Requests and Handling Responses: Send natural language requests just like calling a regular LLM. Fugu internally handles model selection, task delegation, verification, and answer synthesis automatically. The response format received by the user is fully consistent with the OpenAI API, including standard fields like choices and usage. This means existing code can be seamlessly migrated without any modifications.

5. Model Pool Configuration (Optional): In the Sakana AI console, users can customize the model pool. It supports filtering by provider, model name, or capability type to include or exclude specific models from the collaboration pool. For example, you can exclude a provider with unclear data privacy policies or retain only models that excel in mathematical reasoning. Once configured, Fugu will automatically schedule tasks within the selected model pool.

6. Best Practices and Precautions: For complex tasks requiring multi-round deep collaboration, it is recommended to enable the recursive self-calling feature by setting the recursive_depth parameter in the request (default value is 1, can be increased as needed). For cost-sensitive scenarios, start with Fugu Standard and gradually evaluate whether an upgrade to the Ultra version is necessary. Note that excessive recursion depth can significantly increase inference latency; it is recommended to keep it within 3-5 layers depending on task complexity.

4. Pros and Cons Analysis

5. Comparison with Similar Tools

Selection Recommendations: For enterprise users seeking top-tier performance and minimal integration costs, Sakana Fugu is the best choice. Its single API plug-and-play feature enables quick embedding into existing workflows, while dynamic self-learning orchestration eliminates the hassle of manual tuning. It is especially suitable for scenarios requiring diverse task handling and high service resilience, such as financial trading and security analysis. However, if the team has strong AI engineering capabilities and requires highly customized agent behavior, LangChain Agents and CrewAI offer more flexible open-source solutions, ideal for research teams needing deep control over agent behavior.

For cost-sensitive scenarios that do not require complex reasoning, OpenRouter Fusion is a mature alternative. It provides a wide range of model selections and routing strategies but falls short of Fugu in emergent collective intelligence. Note that Fugu currently only offers a closed-source API service; for enterprises requiring local deployment or full control over the underlying architecture, open-source solutions are more suitable.

6. Editor's Take

Sakana Fugu represents a significant leap from engineering to productization in multi-agent systems. Its most notable technical innovation lies in elevating multi-model collaboration from hardcoded pipelines to self-learning, dynamically evolving systems—an approach theoretically validated in the academic papers TRINITY and Conductor, with Fugu being the first system to productize and bring this theory to market. The architecture of "model as orchestrator" is particularly ingenious: by using lightweight heads to predict invocation strategies from hidden states, it achieves efficient collective intelligence emergence at minimal additional computational cost—a technically elegant solution.

From a practical standpoint, Fugu addresses a core pain point in enterprise AI applications: the risk of single-vendor lock-in. Against the backdrop of tightening export controls, the ability to automatically route around unavailable vendor models holds significant strategic importance. At the same time, its cost advantages are undeniable—Fugu Ultra delivers equivalent performance at one-third the price of Opus 4.8, opening the door for budget-constrained enterprises to access cutting-edge AI capabilities.

In terms of target users, Fugu is best suited for two groups: first, enterprise developers seeking peak performance but lacking AI engineering capabilities, who can quickly integrate via a single API; second, industry users sensitive to vendor dependency and requiring high availability assurance, such as those in finance, security, and government sectors. For research-oriented users, Fugu's recursive invocation and dynamic role assignment mechanisms provide an experimental platform for studying the emergence of collective intelligence.

The future development potential is immense. As Sakana AI continues to optimize the TRINITY and Conductor algorithms and expands the model pool, Fugu still has significant room for improvement in reasoning capability, response speed, and cost control. In particular, the deepening of recursive self-invocation mechanisms could give rise to multi-agent systems capable of autonomously planning, executing, and validating complex projects.

7. Application Scenarios

• Software Engineering: Integrate with tools like Codex for real-time code generation, code review, and complex algorithm debugging. Fugu's dynamic role assignment mechanism enables the system to automatically assign a thinker (designing architecture), executor (writing code), and verifier (checking for bugs) for code generation tasks, significantly improving code quality and development efficiency.

• Scientific Research & Competitions: Used for Kaggle competitions, academic paper reproduction, and high-difficulty mathematical and scientific reasoning. The Fugu Ultra version's outstanding performance on benchmarks such as MATH and GPQA makes it a powerful assistant for researchers, with recursive calling capabilities enabling it to handle complex problems requiring multi-step reasoning.

• Security Analysis: Performs exceptionally well on security benchmarks like CTI-REALM, making it suitable for threat analysis and vulnerability research. Fugu's resilient anti-risk architecture ensures stable operation in security-sensitive scenarios, maintaining analysis task continuity even when some models are unavailable.

• Knowledge Work: Conducts literature reviews, patent surveys, long-context reasoning, and in-depth multilingual document analysis. Fugu's multi-model collaboration capability allows it to simultaneously invoke models from different specialized domains, analyzing complex documents from multiple perspectives to enhance analytical depth and breadth.

• Financial Trading: Achieved an average return of +19.43% in a 50-week backtest, outperforming other cutting-edge models. Fugu can simultaneously invoke multiple financial analysis models, reducing the bias of a single model through collective decision-making and improving the robustness of trading strategies.

8. FAQ

Q: What is the difference between Sakana Fugu and directly calling multiple APIs?
A: Directly calling multiple APIs requires developers to design routing strategies, handle model selection, task delegation, and result verification, which typically involves extensive engineering work. Fugu automates these processes, completing all collaborative scheduling through a single API, and dynamically optimizes collaboration strategies based on the TRINITY evolutionary algorithm, often achieving better performance than manual configuration.

Q: What is the core difference between Fugu Ultra and the standard version?
A: Fugu Ultra coordinates a deeper expert pool, usually including more high-performance models and more complex collaboration strategies. On benchmarks requiring deep reasoning, such as MATH and GPQA, the Ultra version significantly outperforms the standard version. The trade-off is higher inference latency and cost. The standard version is better suited for low-latency scenarios like daily coding and conversations.

Q: How does Fugu handle model unavailability?
A: Fugu's underlying model pool has elastic replacement capabilities. When a specific vendor's model becomes unavailable (e.g., service outage, API limits), the system automatically routes requests to other available models, ensuring uninterrupted service. This process is completely transparent to users and requires no manual intervention.

Q: Do I need to worry about data privacy when using Fugu?
A: Fugu supports flexible model pool management, allowing users to independently select or exclude specific vendor models in the console. For sensitive data, processing can be restricted to models compliant with data privacy policies. Additionally, Sakana AI provides encrypted data transmission and processing, but specific data storage policies should be referenced from Sakana AI's privacy policy.

Q: What is Fugu's pricing model?
A: Fugu uses a token-based pricing model, similar to the OpenAI API. The standard and Ultra versions have different pricing tables. According to official data, Fugu Ultra's input price is about one-third of Opus 4.8, and the output price is less than half. For specific pricing, please refer to the latest pricing page on the Sakana AI official website.

Q: Which underlying models does Fugu support?
A: Fugu supports a variety of mainstream models, including but not limited to OpenAI's GPT series, Anthropic's Claude series, Google's Gemini series, and open-source models like LLaMA and Mistral. The specific model pool list is dynamically updated based on Sakana AI's agreements and service status. Users can view the currently available model list in the console.

9. Project Address

• Project Website: https://sakana.ai/fugu/
• Sakana AI Official Blog: https://sakana.ai/blog/