Sakana Fugu

Product Introduction

1. Definition: Sakana Fugu is a multi-agent system delivered as a unified, OpenAI-compatible API model. It functions as a model orchestrator or "meta-model" that dynamically coordinates a pool of diverse, powerful foundation models (like LLMs) to solve complex, multi-step tasks.
2. Core Value Proposition: It provides frontier-level performance without single-vendor dependency. By intelligently routing tasks and synthesizing outputs from a pool of specialized agents, Fugu offers a single API endpoint to access optimized collective intelligence for enhanced reasoning, coding, and agentic workflows.

Main Features

1. Dynamic Model Orchestration: Fugu does not rely on fixed, hand-coded workflows. Instead, it learns to dynamically assemble and coordinate agents from a model pool. Using techniques from its core research (TRINITY and Conductor papers), it assigns roles like "Thinker," "Worker," and "Verifier" across multiple turns to efficiently delegate sub-tasks in coding, math, and reasoning, discovering non-obvious but highly effective collaboration patterns.
2. Unified API & Flexible Model Pool: Access is provided through a single OpenAI-compatible API endpoint, eliminating the complexity of integrating and switching between multiple model providers. Users, particularly on the Fugu tier, can customize the model pool by opting out of specific providers or models to meet data privacy, compliance, or organizational requirements, offering crucial flexibility for enterprise deployment.
3. Two-Tiered Model Offering (Fugu & Fugu Ultra): The system provides two models through the same API. Fugu balances strong performance with low latency for everyday, responsive tasks like chatbots and code review. Fugu Ultra is optimized for maximum answer quality on high-stakes, complex problems, coordinating a deeper pool of experts for tasks like paper reproduction, security analysis, and competitive programming.

Problems Solved

1. Pain Point: Businesses and developers face fragmented AI tooling, requiring management of multiple APIs, vendor lock-in, and suboptimal performance when using a single model for all tasks. They also struggle with the high cognitive overhead of selecting and integrating the right model for each specific sub-task.
2. Target Audience: This includes Software Engineers (for code generation and review), Researchers (for literature/patent analysis and paper reproduction), Data Scientists (for complex Kaggle competitions), Security Engineers (for autonomous security assessments), and Platform Executives (building AI agent products needing stable, orchestrated multi-model intelligence).
3. Use Cases: Essential for multi-step agentic workflows such as end-to-end autonomous security scanning (recon to report), autonomous ML research (iterating on training recipes), solving complex puzzles like blindfold chess or Rubik's cube solver generation, and deep, cross-domain analysis that requires coordinating different expert capabilities sequentially.

Unique Advantages

1. Differentiation: Unlike traditional single-model APIs (e.g., from OpenAI, Anthropic, Google) or basic ensembles, Fugu is a learned orchestration system. It doesn't just call models; it dynamically routes and coordinates them based on learned strategies, often outperforming individual frontier models. The "shoulder-to-shoulder" performance with unreleased models like Fable 5 and Mythos Preview, while avoiding export control risks, is a significant differentiator.
2. Key Innovation: The core innovation is the research-driven coordination framework based on the ICLR 2026 papers "TRINITY" (evolved LLM coordinator) and "Conductor" (RL for natural-language coordination strategies). This allows the system to discover and implement efficient, non-obvious agent collaboration patterns autonomously, rather than relying on human-designed rules.

Frequently Asked Questions (FAQ)

1. How does Sakana Fugu pricing work with multiple agents? You are never charged for multiple models simultaneously. Fugu employs a single blended rate. If multiple agents are active, you pay only one rate—the rate of the top-tier model involved in that request. Fugu Ultra has fixed, per-token pricing.
2. Is my data used to train models, and can I control which models Fugu uses? Usage data can improve Fugu, but you can opt out of data sharing at any time. For the Fugu model, you can control the model pool by opting out of specific providers via the console to meet data and compliance needs. Fugu Ultra's pool is fixed to ensure maximum performance.
3. What is the main difference between Fugu and Fugu Ultra? Fugu is the default model, optimized for a balance of performance and low latency, ideal for daily coding, chatbots, and interactive work. Fugu Ultra prioritizes maximum answer quality on complex, multi-step problems, using a larger expert pool at the cost of higher latency and token cost, suited for research, analysis, and competitions.