Sakana Fugu

What is Sakana Fugu?

Sakana Fugu is a groundbreaking multi-agent AI platform that delivers frontier-level performance by dynamically orchestrating the world's best models to tackle complex, multi-step tasks. Rather than relying on a single AI model, Fugu intelligently coordinates a diverse pool of specialized models through learned collaboration patterns, achieving superior results without single-vendor dependency.

Core Capabilities & Models

Fugu is available in three specialized models, all accessible through a single OpenAI-compatible API:

• Fugu: Balanced performance and low latency for everyday coding, code review, and responsive chatbot services
• Fugu Ultra: Optimized for maximum answer quality on high-stakes problems like Kaggle competitions, paper reproduction, cybersecurity analysis, and patent investigations
• Fugu Cyber: Specialized for cybersecurity reasoning tasks including security analysis, vulnerability research, and threat investigation

Who It's For

Fugu is designed for engineers, researchers, data scientists, and organizations requiring frontier AI capabilities for quality-critical workflows. It's particularly valuable for teams working on coding, scientific reasoning, cybersecurity, and complex multi-step automation where accuracy and reliability are paramount.

Key Differentiators

• Dynamic Model Orchestration: Instead of prescribing fixed workflows, Fugu learns to dynamically assemble and coordinate agents through non-obvious but highly efficient collaboration patterns
• Frontier Performance: Surpasses publicly accessible frontier models on rigorous benchmarks including SWE Bench Pro (73.7%), LiveCodeBench (93.2%), and CyberGym (86.9%)
• Flexible Agent Selection: Control which models participate in Fugu's pool to meet data privacy, compliance, and organizational requirements
• Single API Integration: Access coordinated multi-model intelligence through one endpoint, reducing API complexity while improving cost-performance
• Export Control Freedom: Delivers frontier capability without the risk of export controls

Research Foundation

Fugu is grounded in two ICLR 2026 papers—TRINITY and the Conductor—demonstrating how systems can learn to assemble, route, and coordinate expert agents for each task through evolved coordination and reinforcement learning rather than hand-designed workflows.

Proven Results

Fugu has demonstrated state-of-the-art performance across coding (LiveCodeBench Pro: 90.8%), reasoning (Humanity's Last Exam: 50.0%), scientific tasks (GPQA-D: 95.5%), and cybersecurity (CTI-REALM: 72.1%), with early users successfully applying it to autonomous ML research, classical Japanese text analysis, Rubik's Cube solving, CAD generation, and time-series trading.