AI Tool

Fugu Review

Name: Fugu
Availability: OnlineOnly
Author: Stork.AI

Fugu is a multi-agent orchestration system functioning as a conductor LLM, trained to dynamically route incoming tasks to the optimal agent from a swappable pool of other LLMs.

shipped Jun 23, 2026freemium

Domain rating73Monthly visits12K/moAI-readablepartial

Why it matters

1Launched on June 22, 2026, by Sakana AI, offering an OpenAI-compatible API.

2Fugu Ultra achieved 73.7% on SWE-Bench Pro and 93.2% on LiveCodeBench.

3Offers freemium access with subscription plans starting at $20/month.

4Fugu Ultra API input tokens are priced at $0.005 per 1k tokens, output at $0.03 per 1k tokens.

Stork’s verdict on Fugu

Fugu excels at complex coding by orchestrating diverse LLMs, but its Ultra variant's black box selection can be slow and expensive.

Fugu reviewed by Stork AI · stork.ai/en/fugu

Specs

API Docs

View Documentation →

API Available

Yes, public API

overview

What is Fugu?

Fugu is a multi-agent orchestration system developed by Sakana AI that enables corporations and engineering teams to dynamically route incoming tasks to the optimal LLM agent. It functions as a conductor LLM, abstracting the complexity of managing multiple model providers through a single, OpenAI-compatible API. Launched on June 22, 2026, Fugu is designed to leverage the collective intelligence of multiple large language models (LLMs) to provide high-performance AI capabilities while mitigating risks associated with single-vendor dependency.

Fugu operates as a 'learned orchestrator,' an AI model trained to receive a request, intelligently decide whether to handle it directly or delegate to specialist models within its agent pool, manage verification and synthesis of responses, and then return a single, consolidated answer. The system is offered in two primary variants:

Fugu: Optimized for speed and lower cost, this variant is suitable for everyday, latency-sensitive tasks such as coding assistance, code review, chatbots, internal assistants, and interactive services. Users can opt out specific agents from its pool for data, privacy, or compliance reasons.
Fugu Ultra: Tuned for maximum answer quality on complex, multi-step, and high-stakes problems. Its use cases include paper reproduction, cybersecurity analysis, Kaggle-style data science, patent investigations, deep technical research, and complex code review.

features

Key Features of Fugu

Fugu, developed by Sakana AI, incorporates several key features designed to optimize multi-agent AI system performance and flexibility.

Multi-agent orchestration system for dynamic task routing.
Functions as a conductor LLM to manage complex requests.
Dynamically routes incoming tasks to the optimal agent from a swappable pool of LLMs.
Provides an OpenAI-compatible API for broad integration.
Offers two variants: Fugu for speed/cost efficiency and Fugu Ultra for maximum answer quality on complex tasks.
Allows users to opt out specific agents from its pool for data, privacy, or compliance reasons.
Supports recursive self-improvement and adaptive foundation models.

use cases

Who Should Use Fugu?

Fugu is designed for organizations and teams requiring advanced AI capabilities with a focus on efficiency, adaptability, and data governance, particularly those seeking to mitigate single-vendor AI dependency.

Corporations & Financial Institutions: For complex, high-stakes problems like cybersecurity analysis, patent investigations, and deep technical research.
Japanese Enterprises & Organizations with Strict Data Governance: To leverage multi-agent AI while maintaining control over data and compliance.
Engineering Teams & Data Science Units: For coding assistance, code review, paper reproduction, and Kaggle-style data science challenges.
Think Tanks & Research Institutions: For deep technical research and complex multi-step problem-solving.

how to use

How to Use Fugu

Fugu is accessible via its OpenAI-compatible API and the Sakana console, allowing users to integrate its multi-agent orchestration capabilities into their applications and workflows.

1Access the Sakana console at console.sakana.ai to manage subscriptions and API keys.
2Utilize the OpenAI-compatible API for Fugu and Fugu Ultra by integrating with existing systems.
3Select between Fugu (model ID: fugu-20260615) for speed-optimized tasks or Fugu Ultra (model ID: fugu-ultra-20260615) for quality-focused complex problems.
4Configure agent pool preferences, including opting out specific agents for privacy or compliance reasons.
5Submit complex, multi-step prompts to Fugu Ultra for tasks like code review or research analysis.

pricing

Fugu Pricing & Plans

Fugu offers a freemium model with both subscription plans and pay-as-you-go API rates, accessible through the Sakana console at console.sakana.ai. A launch offer through July 2026 provides a second month free for new subscribers to subscription plans.

Freemium: Basic access for initial exploration.
Standard Plan: $20/month for baseline usage, suitable for occasional API calls and personal experiments.
Pro Plan: $100/month for 10x the Standard usage, designed for regular coding and research sessions.
Max Plan: $200/month for 30x the Standard usage, intended for heavy, long-running workloads.
Fugu Ultra API (Pay-as-you-go): Input tokens are $0.005 per 1k tokens, and output tokens are $0.03 per 1k tokens (model ID: fugu-ultra-20260615).
Fugu Ultra High-Context API (Pay-as-you-go): Input tokens are $0.01 per 1k tokens, and output tokens are $0.045 per 1k tokens.

Pros

+Achieves high performance on complex, multi-step coding tasks, with Fugu Ultra identifying significantly more issues than other tools in beta tests.
+Mitigates risks associated with single-vendor AI dependency by orchestrating a pool of diverse LLMs.
+Demonstrates strong benchmark results, outperforming competitors like Claude Opus 4.8 and GPT-5.5 on SWE-Bench Pro (73.7%) and LiveCodeBench (93.2%).
+Offers an OpenAI-compatible API, simplifying integration into existing developer workflows.
+Provides flexibility with two variants (Fugu and Fugu Ultra) optimized for different use cases and cost/quality priorities.
+Allows users to opt out specific agents from its pool for enhanced data privacy and compliance.

Cons

−Fugu Ultra can be expensive and consume tokens rapidly; the $200/month Max plan could be exhausted in under three hours of heavy use per week.
−Users have reported slow processing speeds for Fugu Ultra, leading to stress and potentially making manual guidance of other frontier models more efficient for complex tasks.
−The internal model selection and orchestration process is a 'black box,' raising concerns about accountability and transparency.
−Fugu Ultra trailed Anthropic's Fable 5 on Humanity's Last Exam (50.0% vs. 53.3%) and SWE-Bench Pro (73.7% vs. 86.0%).
−Critics suggest Fugu could be a 'de facto downgrade' if its agent pool primarily consists of lower-tier open models, excluding the most advanced, regulated AI systems.

Similar Tools

Fugu vs Competitors

Sakana AI positions Fugu as a 'genuinely different bet' against monolithic frontier models, aiming to achieve or exceed frontier performance through learned orchestration of existing LLMs. Fugu Ultra has demonstrated strong benchmark results, including 73.7% on SWE-Bench Pro, 93.2% on LiveCodeBench, 95.5% on GPQA Diamond, and 82.1% on TerminalBench 2.1.

Microsoft AutoGenOn Stork

Its chat-centric orchestration model enables asynchronous communication and sophisticated negotiation among various agents, including humans, LLM-agents, and code executors.

Like Fugu, AutoGen orchestrates multiple LLMs and tools into cooperative workflows, but it particularly emphasizes conversational multi-agent systems and a unified message-passing interface for scalability. Its SDK is free, aligning with Fugu's freemium model.

CrewAIOn Stork

It provides a role-based multi-agent framework where specialized AI agents collaborate on complex tasks, featuring sophisticated memory management and checkpointing.

CrewAI, similar to Fugu, focuses on orchestrating multiple AI agents from a pool, but it specifically structures agents with distinct roles and a defined process for task handoffs and communication. It is an open-source framework, aligning with Fugu's freemium approach.

LangGraphOn Stork Compare

It is a framework for building stateful, multi-actor applications with LLMs, offering granular control over an agent's thought process for reliable and complex AI agents.

While Fugu focuses on dynamic routing to optimal LLMs, LangGraph provides a more expressive framework for building complex, cyclical, and stateful multi-actor agentic workflows, often building on LangChain's foundational components. It is an MIT-licensed open-source library and free to use, similar to Fugu's freemium model.

RouteLLMOn Stork

It is an open-source framework specifically designed for dynamically selecting the most suitable LLM for each query based on complexity and performance requirements to optimize costs.

RouteLLM directly competes with Fugu's core functionality of dynamically routing incoming tasks to the optimal agent by focusing on cost-effective LLM routing. Unlike Fugu's broader multi-agent orchestration, RouteLLM is more specialized in the routing decision itself, and it is open-source.