Sakana AI Introduces Fugu Orchestration Model for Advanced LLM Task Routing

Sakana AI has officially launched Fugu, an innovative multi-agent orchestration system that streamlines how complex tasks are handled by large language models (LLMs). This new system functions as a single endpoint, intelligently deciding whether to resolve a task directly or to assemble and coordinate a team of expert models from a diverse pool. The Fugu system aims to simplify the integration and management of multiple LLMs, presenting a unified interface while managing internal complexities like model selection, delegation, verification, and synthesis.

The introduction of Fugu, alongside its more powerful variant Fugu Ultra, marks a significant step towards more adaptable and robust AI architectures. Sakana AI highlights that this approach mitigates risks associated with single-vendor dependencies, allowing developers to maintain continuity even if access to specific models is restricted—a concern underscored by recent export controls on models like Anthropic’s Fable and Mythos. New models can be integrated into the Fugu pool over time, ensuring future-proofing and flexibility.

How Fugu Works

Fugu itself is an advanced language model trained specifically to call upon and manage other LLMs within its agent pool, which can even include recursive instances of itself. Unlike traditional hard-coded workflows, Fugu learns dynamically how to coordinate tasks, delegate responsibilities, and facilitate communication among agents. This learned orchestration capability is built upon research presented in two ICLR 2026 papers, Trinity and Conductor.

Trinity employs a lightweight, evolved coordinator that adaptively assigns roles such as Thinker, Worker, or Verifier across multiple turns to delegate work. Conductor, on the other hand, is trained using reinforcement learning to discover natural-language coordination strategies and generate focused prompts for diverse LLM pools. Together, these foundational concepts enable Fugu to replace manual, hand-designed workflows with an intelligent, adaptive system.

Performance Benchmarks and API Compatibility

Sakana AI reports that Fugu and Fugu Ultra demonstrate leading performance across various benchmarks. When compared against the foundation models it orchestrates, Fugu Ultra topped four coding benchmarks, CharXiv Reasoning, and Humanity’s Last Exam. It also tied with regular Fugu on GPQA-D. Regular Fugu led in SciCode, τ³ Banking, and Long Context Reasoning. Notably, the Fugu models perform comparably to high-performing models like Anthropic’s Fable 5 and Mythos Preview, even though these are not part of Fugu’s accessible pool.

For developers, Fugu offers an OpenAI-compatible API, meaning existing client integrations require no SDK migration. Developers can simply point their existing OpenAI clients to a console-provided endpoint. This ease of integration is a key factor in accelerating adoption, as seen in a beta program involving nearly 500 early users, with published examples often showcasing its capability in handling long, multi-step tasks. Token usage and costs are reported per request, allowing for real-time spend monitoring.

Key facts

Feature	Description
Model Name	Sakana Fugu (and Fugu Ultra)
Function	Multi-agent orchestration system for routing tasks across a pool of frontier LLMs
Core Benefit	Reduces single-vendor dependency, enhances performance, simplifies complex LLM interactions
API	OpenAI-compatible API

Impact for Indian AI Developers and Businesses

For the Indian technology and startup ecosystem, the launch of Sakana Fugu presents significant opportunities. As India increasingly focuses on AI innovation and digital transformation, tools that offer flexibility, reduce vendor lock-in, and improve the efficiency of AI model deployment are crucial. Indian startups and developers striving to build advanced AI applications can leverage Fugu to tap into a broader range of LLMs without the overhead of managing individual integrations or being limited by a single provider’s offerings. This could accelerate the development of sophisticated AI solutions in areas such as coding, reasoning, and multi-step agentic tasks, aligning with the goals of initiatives like the IndiaAI Mission.

Source: MarkTechPost, https://www.marktechpost.com/2026/06/22/sakana-ai-launches-sakana-fugu-an-orchestration-model-that-routes-tasks-across-a-swappable-pool-of-frontier-llms/